Abstract
Refurbishing laptops extends their lives while contributing to reducing electronic waste, which promotes building a sustainable future. To this end, the Danish Technological Institute (DTI) focuses on the research and development of several applications, including laptop refurbishing. This has several steps, including cleaning, which involves identifying and removing stickers from laptop surfaces. DTI trained six sticker detection models (SDMs) based on open-source object detection models to identify such stickers precisely so these stickers can be removed automatically. However, given the diversity in types of stickers (e.g., shapes, colors, locations), identification of the stickers is highly uncertain, thereby requiring explicit quantification of uncertainty associated with the identified stickers. Such uncertainty quantification can help reduce risks in removing stickers, which, for example, could otherwise result in damaging laptop surfaces. For uncertainty quantification, we adopted the Monte Carlo Dropout method to evaluate the six SDMs from DTI using three datasets: the original image dataset from DTI and two datasets generated with vision language models, i.e., DALL-E-3 and Stable Diffusion-3. In addition, we presented novel robustness metrics concerning detection accuracy and uncertainty to assess the robustness of the SDMs based on adversarial datasets generated from the three datasets using a dense adversary method. Our evaluation results show that different SDMs perform differently regarding different metrics. Based on the results, we provide SDM selection guidelines and lessons learned from various perspectives.
Title:
BFA-YOLO: Balanced multiscale object detection network for multi-view building facade attachments detection
Abstract
Detection of building facade attachments such as doors, windows, balconies, air conditioner units, billboards, and glass curtain walls plays a pivotal role in numerous applications. Building facade attachments detection aids in vbuilding information modeling (BIM) construction and meeting Level of Detail 3 (LOD3) standards. Yet, it faces challenges like uneven object distribution, small object detection difficulty, and background interference. To counter these, we propose BFA-YOLO, a model for detecting facade attachments in multi-view images. BFA-YOLO incorporates three novel innovations: the Feature Balanced Spindle Module (FBSM) for addressing uneven distribution, the Target Dynamic Alignment Task Detection Head (TDATH) aimed at improving small object detection, and the Position Memory Enhanced Self-Attention Mechanism (PMESA) to combat background interference, with each component specifically designed to solve its corresponding challenge. Detection efficacy of deep network models deeply depends on the dataset's characteristics. Existing open source datasets related to building facades are limited by their single perspective, small image pool, and incomplete category coverage. We propose a novel method for building facade attachments detection dataset construction and construct the BFA-3D dataset for facade attachments detection. The BFA-3D dataset features multi-view, accurate labels, diverse categories, and detailed classification. BFA-YOLO surpasses YOLOv8 by 1.8% and 2.9% in mAP@0.5 on the multi-view BFA-3D and street-view Facade-WHU datasets, respectively. These results underscore BFA-YOLO's superior performance in detecting facade attachments.
Title:
D4: Text-guided diffusion model-based domain adaptive data augmentation for vineyard shoot detection
Abstract
In an agricultural field, plant phenotyping using object detection models is gaining attention. However, collecting the training data necessary to create generic and high-precision models is extremely challenging due to the difficulty of annotation and the diversity of domains. Furthermore, it is difficult to transfer training data across different crops, and although machine learning models effective for specific environments, conditions, or crops have been developed, they cannot be widely applied in actual fields. In this study, we propose a generative data augmentation method (D4) for vineyard shoot detection. D4 uses a pre-trained text-guided diffusion model based on a large number of original images culled from video data collected by unmanned ground vehicles or other means, and a small number of annotated datasets. The proposed method generates new annotated images with background information adapted to the target domain while retaining annotation information necessary for object detection. In addition, D4 overcomes the lack of training data in agriculture, including the difficulty of annotation and diversity of domains. We confirmed that this generative data augmentation method improved the mean average precision by up to 28.65% for the BBox detection task and the average precision by up to 13.73% for the keypoint detection task for vineyard shoot detection. Our generative data augmentation method D4 is expected to simultaneously solve the cost and domain diversity issues of training data generation in agriculture and improve the generalization performance of detection models.
Title:
UniDet3D: Multi-dataset Indoor 3D Object Detection
Authors: Maksim Kolodiazhnyi, Anna Vorontsova, Matvey Skripkin, Danila Rukhovich, Anton Konushin
Subjects: Subjects:
Computer Vision and Pattern Recognition (cs.CV)
Abstract
Growing customer demand for smart solutions in robotics and augmented reality has attracted considerable attention to 3D object detection from point clouds. Yet, existing indoor datasets taken individually are too small and insufficiently diverse to train a powerful and general 3D object detection model. In the meantime, more general approaches utilizing foundation models are still inferior in quality to those based on supervised training for a specific task. In this work, we propose \ours{}, a simple yet effective 3D object detection model, which is trained on a mixture of indoor datasets and is capable of working in various indoor environments. By unifying different label spaces, \ours{} enables learning a strong representation across multiple datasets through a supervised joint training scheme. The proposed network architecture is built upon a vanilla transformer encoder, making it easy to run, customize and extend the prediction pipeline for practical use. Extensive experiments demonstrate that \ours{} obtains significant gains over existing 3D object detection methods in 6 indoor benchmarks: ScanNet (+1.1 mAP50), ARKitScenes (+19.4 mAP25), S3DIS (+9.1 mAP50), MultiScan (+9.3 mAP50), 3RScan (+3.2 mAP50), and ScanNet++ (+2.7 mAP50). Code is available at this https URL .
Title:
Future Does Matter: Boosting 3D Object Detection with Temporal Motion Estimation in Point Cloud Sequences
Authors: Rui Yu, Runkai Zhao, Cong Nie, Heng Wang, HuaiCheng Yan, Meng Wang
Subjects: Subjects:
Computer Vision and Pattern Recognition (cs.CV)
Abstract
Accurate and robust LiDAR 3D object detection is essential for comprehensive scene understanding in autonomous driving. Despite its importance, LiDAR detection performance is limited by inherent constraints of point cloud data, particularly under conditions of extended distances and occlusions. Recently, temporal aggregation has been proven to significantly enhance detection accuracy by fusing multi-frame viewpoint information and enriching the spatial representation of objects. In this work, we introduce a novel LiDAR 3D object detection framework, namely LiSTM, to facilitate spatial-temporal feature learning with cross-frame motion forecasting information. We aim to improve the spatial-temporal interpretation capabilities of the LiDAR detector by incorporating a dynamic prior, generated from a non-learnable motion estimation model. Specifically, Motion-Guided Feature Aggregation (MGFA) is proposed to utilize the object trajectory from previous and future motion states to model spatial-temporal correlations into gaussian heatmap over a driving sequence. This motion-based heatmap then guides the temporal feature fusion, enriching the proposed object features. Moreover, we design a Dual Correlation Weighting Module (DCWM) that effectively facilitates the interaction between past and prospective frames through scene- and channel-wise feature abstraction. In the end, a cascade cross-attention-based decoder is employed to refine the 3D prediction. We have conducted experiments on the Waymo and nuScenes datasets to demonstrate that the proposed framework achieves superior 3D detection performance with effective spatial-temporal feature learning.
Keyword: transformer
Title:
MVTN: A Multiscale Video Transformer Network for Hand Gesture Recognition
Abstract
In this paper, we introduce a novel Multiscale Video Transformer Network (MVTN) for dynamic hand gesture recognition, since multiscale features can extract features with variable size, pose, and shape of hand which is a challenge in hand gesture recognition. The proposed model incorporates a multiscale feature hierarchy to capture diverse levels of detail and context within hand gestures which enhances the model's ability. This multiscale hierarchy is obtained by extracting different dimensions of attention in different transformer stages with initial stages to model high-resolution features and later stages to model low-resolution features. Our approach also leverages multimodal data, utilizing depth maps, infrared data, and surface normals along with RGB images from NVGesture and Briareo datasets. Experiments show that the proposed MVTN achieves state-of-the-art results with less computational complexity and parameters. The source code is available at this https URL.
Title:
On-board Satellite Image Classification for Earth Observation: A Comparative Study of Pre-Trained Vision Transformer Models
Authors: Thanh-Dung Le, Vu Nguyen Ha, Ti Ti Nguyen, Geoffrey Eappen, Prabhu Thiruvasagam, Luis M. Garces-Socarras, Hong-fu Chou, Jorge L. Gonzalez-Rios, Juan Carlos Merlano-Duncan, Symeon Chatzinotas
Subjects: Subjects:
Computer Vision and Pattern Recognition (cs.CV); Signal Processing (eess.SP)
Abstract
Remote sensing image classification is a critical component of Earth observation (EO) systems, traditionally dominated by convolutional neural networks (CNNs) and other deep learning techniques. However, the advent of Transformer-based architectures and large-scale pre-trained models has significantly shifted, offering enhanced performance and efficiency. This study focuses on identifying the most effective pre-trained model for land use classification in onboard satellite processing, emphasizing achieving high accuracy, computational efficiency, and robustness against noisy data conditions commonly encountered during satellite-based inference. Through extensive experimentation, we compared traditional CNN-based models, ResNet-based models, and various pre-trained vision Transformer models. Our findings demonstrate that pre-trained Transformer models, particularly MobileViTV2 and EfficientViT-M2, outperform models trained from scratch in accuracy and efficiency. These models achieve high performance with reduced computational requirements and exhibit greater resilience during inference under noisy conditions. While MobileViTV2 excelled on clean validation data, EfficientViT-M2 proved more robust when handling noise, making it the most suitable model for onboard satellite Earth observation tasks. In conclusion, EfficientViT-M2 is the optimal choice for reliable and efficient remote sensing image classification in satellite operations, achieving 98.76\% accuracy, precision, and recall. Specifically, EfficientViT-M2 delivered the highest performance across all metrics, excelled in training efficiency (1,000s) and inference time (10s), and demonstrated greater robustness (overall robustness score at 0.79).
Title:
CACER: Clinical Concept Annotations for Cancer Events and Relations
Authors: Yujuan Fu, Giridhar Kaushik Ramachandran, Ahmad Halwani, Bridget T. McInnes, Fei Xia, Kevin Lybarger, Meliha Yetisgen, Özlem Uzuner
Subjects: Subjects:
Computation and Language (cs.CL)
Abstract
Clinical notes contain unstructured representations of patient histories, including the relationships between medical problems and prescription drugs. To investigate the relationship between cancer drugs and their associated symptom burden, we extract structured, semantic representations of medical problem and drug information from the clinical narratives of oncology notes. We present Clinical Concept Annotations for Cancer Events and Relations (CACER), a novel corpus with fine-grained annotations for over 48,000 medical problems and drug events and 10,000 drug-problem and problem-problem relations. Leveraging CACER, we develop and evaluate transformer-based information extraction (IE) models such as BERT, Flan-T5, Llama3, and GPT-4 using fine-tuning and in-context learning (ICL). In event extraction, the fine-tuned BERT and Llama3 models achieved the highest performance at 88.2-88.0 F1, which is comparable to the inter-annotator agreement (IAA) of 88.4 F1. In relation extraction, the fine-tuned BERT, Flan-T5, and Llama3 achieved the highest performance at 61.8-65.3 F1. GPT-4 with ICL achieved the worst performance across both tasks. The fine-tuned models significantly outperformed GPT-4 in ICL, highlighting the importance of annotated training data and model optimization. Furthermore, the BERT models performed similarly to Llama3. For our task, LLMs offer no performance advantage over the smaller BERT models. The results emphasize the need for annotated training data to optimize models. Multiple fine-tuned transformer models achieved performance comparable to IAA for several extraction tasks.
Title:
FODA-PG for Enhanced Medical Imaging Narrative Generation: Adaptive Differentiation of Normal and Abnormal Attributes
Authors: Kai Shu, Yuzhuo Jia, Ziyang Zhang, Jiechao Gao
Abstract
Automatic Medical Imaging Narrative generation aims to alleviate the workload of radiologists by producing accurate clinical descriptions directly from radiological images. However, the subtle visual nuances and domain-specific terminology in medical images pose significant challenges compared to generic image captioning tasks. Existing approaches often neglect the vital distinction between normal and abnormal findings, leading to suboptimal performance. In this work, we propose FODA-PG, a novel Fine-grained Organ-Disease Adaptive Partitioning Graph framework that addresses these limitations through domain-adaptive learning. FODA-PG constructs a granular graphical representation of radiological findings by separating disease-related attributes into distinct "disease-specific" and "disease-free" categories based on their clinical significance and location. This adaptive partitioning enables our model to capture the nuanced differences between normal and pathological states, mitigating the impact of data biases. By integrating this fine-grained semantic knowledge into a powerful transformer-based architecture and providing rigorous mathematical justifications for its effectiveness, FODA-PG generates precise and clinically coherent reports with enhanced generalization capabilities. Extensive experiments on the IU-Xray and MIMIC-CXR benchmarks demonstrate the superiority of our approach over state-of-the-art methods, highlighting the importance of domain adaptation in medical report generation.
Title:
Qihoo-T2X: An Efficiency-Focused Diffusion Transformer via Proxy Tokens for Text-to-Any-Task
Abstract
The global self-attention mechanism in diffusion transformers involves redundant computation due to the sparse and redundant nature of visual information, and the attention map of tokens within a spatial window shows significant similarity. To address this redundancy, we propose the Proxy Token Diffusion Transformer (PT-DiT), which employs sparse representative token attention (where the number of representative tokens is much smaller than the total number of tokens) to model global visual information efficiently. Specifically, in each transformer block, we randomly sample one token from each spatial-temporal window to serve as a proxy token for that region. The global semantics are captured through the self-attention of these proxy tokens and then injected into all latent tokens via cross-attention. Simultaneously, we introduce window and shift window attention to address the limitations in detail modeling caused by the sparse attention mechanism. Building on the well-designed PT-DiT, we further develop the Qihoo-T2X family, which includes a variety of models for T2I, T2V, and T2MV tasks. Experimental results show that PT-DiT achieves competitive performance while reducing the computational complexity in both image and video generation tasks (e.g., a 48% reduction compared to DiT and a 35% reduction compared to Pixart-alpha). Our source code is available at this https URL.
Title:
SDformerFlow: Spatiotemporal swin spikeformer for event-based optical flow estimation
Abstract
Event cameras generate asynchronous and sparse event streams capturing changes in light intensity. They offer significant advantages over conventional frame-based cameras, such as a higher dynamic range and an extremely faster data rate, making them particularly useful in scenarios involving fast motion or challenging lighting conditions. Spiking neural networks (SNNs) share similar asynchronous and sparse characteristics and are well-suited for processing data from event cameras. Inspired by the potential of transformers and spike-driven transformers (spikeformers) in other computer vision tasks, we propose two solutions for fast and robust optical flow estimation for event cameras: STTFlowNet and SDformerFlow. STTFlowNet adopts a U-shaped artificial neural network (ANN) architecture with spatiotemporal shifted window self-attention (swin) transformer encoders, while SDformerFlow presents its fully spiking counterpart, incorporating swin spikeformer encoders. Furthermore, we present two variants of the spiking version with different neuron models. Our work is the first to make use of spikeformers for dense optical flow estimation. We conduct end-to-end training for all models using supervised learning. Our results yield state-of-the-art performance among SNN-based event optical flow methods on both the DSEC and MVSEC datasets, and show significant reduction in power consumption compared to the equivalent ANNs.
Title:
UNIT: Unifying Image and Text Recognition in One Vision Encoder
Authors: Yi Zhu, Yanpeng Zhou, Chunwei Wang, Yang Cao, Jianhua Han, Lu Hou, Hang Xu
Subjects: Subjects:
Computer Vision and Pattern Recognition (cs.CV)
Abstract
Currently, vision encoder models like Vision Transformers (ViTs) typically excel at image recognition tasks but cannot simultaneously support text recognition like human visual recognition. To address this limitation, we propose UNIT, a novel training framework aimed at UNifying Image and Text recognition within a single model. Starting with a vision encoder pre-trained with image recognition tasks, UNIT introduces a lightweight language decoder for predicting text outputs and a lightweight vision decoder to prevent catastrophic forgetting of the original image encoding capabilities. The training process comprises two stages: intra-scale pretraining and inter-scale finetuning. During intra-scale pretraining, UNIT learns unified representations from multi-scale inputs, where images and documents are at their commonly used resolution, to enable fundamental recognition capability. In the inter-scale finetuning stage, the model introduces scale-exchanged data, featuring images and documents at resolutions different from the most commonly used ones, to enhance its scale robustness. Notably, UNIT retains the original vision encoder architecture, making it cost-free in terms of inference and deployment. Experiments across multiple benchmarks confirm that our method significantly outperforms existing methods on document-related tasks (e.g., OCR and DocQA) while maintaining the performances on natural images, demonstrating its ability to substantially enhance text recognition without compromising its core image recognition capabilities.
Title:
Prompt-based Personality Profiling: Reinforcement Learning for Relevance Filtering
Authors: Jan Hofmann, Cornelia Sindermann, Roman Klinger
Subjects: Subjects:
Computation and Language (cs.CL)
Abstract
Author profiling is the task of inferring characteristics about individuals by analyzing content they share. Supervised machine learning still dominates automatic systems that perform this task, despite the popularity of prompting large language models to address natural language understanding tasks. One reason is that the classification instances consist of large amounts of posts, potentially a whole user profile, which may exceed the input length of Transformers. Even if a model can use a large context window, the entirety of posts makes the application of API-accessed black box systems costly and slow, next to issues which come with such "needle-in-the-haystack" tasks. To mitigate this limitation, we propose a new method for author profiling which aims at distinguishing relevant from irrelevant content first, followed by the actual user profiling only with relevant data. To circumvent the need for relevance-annotated data, we optimize this relevance filter via reinforcement learning with a reward function that utilizes the zero-shot capabilities of large language models. We evaluate our method for Big Five personality trait prediction on two Twitter corpora. On publicly available real-world data with a skewed label distribution, our method shows similar efficacy to using all posts in a user profile, but with a substantially shorter context. An evaluation on a version of these data balanced with artificial posts shows that the filtering to relevant posts leads to a significantly improved accuracy of the predictions.
Title:
Context is the Key: Backdoor Attacks for In-Context Learning with Vision Transformers
Authors: Gorka Abad, Stjepan Picek, Lorenzo Cavallaro, Aitor Urbieta
Subjects: Subjects:
Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
Abstract
Due to the high cost of training, large model (LM) practitioners commonly use pretrained models downloaded from untrusted sources, which could lead to owning compromised models. In-context learning is the ability of LMs to perform multiple tasks depending on the prompt or context. This can enable new attacks, such as backdoor attacks with dynamic behavior depending on how models are prompted. In this paper, we leverage the ability of vision transformers (ViTs) to perform different tasks depending on the prompts. Then, through data poisoning, we investigate two new threats: i) task-specific backdoors where the attacker chooses a target task to attack, and only the selected task is compromised at test time under the presence of the trigger. At the same time, any other task is not affected, even if prompted with the trigger. We succeeded in attacking every tested model, achieving up to 89.90\% degradation on the target task. ii) We generalize the attack, allowing the backdoor to affect \emph{any} task, even tasks unseen during the training phase. Our attack was successful on every tested model, achieving a maximum of $13\times$ degradation. Finally, we investigate the robustness of prompts and fine-tuning as techniques for removing the backdoors from the model. We found that these methods fall short and, in the best case, reduce the degradation from 89.90\% to 73.46\%.
Title:
GALLa: Graph Aligned Large Language Models for Improved Source Code Understanding
Authors: Ziyin Zhang, Hang Yu, Shijie Li, Peng Di, Jianguo Li, Rui Wang
Subjects: Subjects:
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Abstract
Programming languages possess rich semantic information such as data flow that is represented by graphs and not available from the surface form of source code. Recent code language models have scaled to billions of parameters, but model source code solely as text tokens while ignoring any other structural information. Conversely, models that do encode structural information of code make modifications to the Transformer architecture, limiting their scale and compatibility with pretrained LLMs. In this work, we take the best of both worlds with GALLa - Graph Aligned Large Language Model. GALLa utilizes graph neural networks and cross-modal alignment technologies to inject the structural information of code into LLMs as an auxiliary task during finetuning. This framework is both model-agnostic and task-agnostic, as it can be applied to any code LLM for any code downstream task, and requires the structural graph data only at training time from a corpus unrelated to the finetuning data, while incurring no cost at inference time over the baseline LLM. Experiments on five code tasks with four different baseline LLMs ranging in size from 350M to 8B validate the effectiveness of GALLa, demonstrating consistent improvement over the baseline, even for powerful models such as LLaMA3.
Title:
Residual Stream Analysis with Multi-Layer SAEs
Authors: Tim Lawson, Lucy Farnik, Conor Houghton, Laurence Aitchison
Subjects: Subjects:
Machine Learning (cs.LG); Computation and Language (cs.CL)
Abstract
Sparse autoencoders (SAEs) are a promising approach to interpreting the internal representations of transformer language models. However, standard SAEs are trained separately on each transformer layer, making it difficult to use them to study how information flows across layers. To solve this problem, we introduce the multi-layer SAE (MLSAE): a single SAE trained on the residual stream activation vectors from every transformer layer simultaneously. The residual stream is usually understood as preserving information across layers, so we expected to, and did, find individual SAE features that are active at multiple layers. Interestingly, while a single SAE feature is active at different layers for different prompts, for a single prompt, we find that a single feature is far more likely to be active at a single layer. For larger underlying models, we find that the cosine similarities between adjacent layers in the residual stream are higher, so we expect more features to be active at multiple layers. These results show that MLSAEs are a promising method to study information flow in transformers. We release our code to train and analyze MLSAEs at this https URL.
Title:
GST: Precise 3D Human Body from a Single Image with Gaussian Splatting Transformers
Authors: Lorenza Prospero, Abdullah Hamdi, Joao F. Henriques, Christian Rupprecht
Abstract
Reconstructing realistic 3D human models from monocular images has significant applications in creative industries, human-computer interfaces, and healthcare. We base our work on 3D Gaussian Splatting (3DGS), a scene representation composed of a mixture of Gaussians. Predicting such mixtures for a human from a single input image is challenging, as it is a non-uniform density (with a many-to-one relationship with input pixels) with strict physical constraints. At the same time, it needs to be flexible to accommodate a variety of clothes and poses. Our key observation is that the vertices of standardized human meshes (such as SMPL) can provide an adequate density and approximate initial position for Gaussians. We can then train a transformer model to jointly predict comparatively small adjustments to these positions, as well as the other Gaussians' attributes and the SMPL parameters. We show empirically that this combination (using only multi-view supervision) can achieve fast inference of 3D human models from a single image without test-time optimization, expensive diffusion models, or 3D points supervision. We also show that it can improve 3D pose estimation by better fitting human models that account for clothes and other variations. The code is available on the project website this https URL .
Title:
Introducing Gating and Context into Temporal Action Detection
Authors: Aglind Reka, Diana Laura Borza, Dominick Reilly, Michal Balazia, Francois Bremond
Subjects: Subjects:
Computer Vision and Pattern Recognition (cs.CV)
Abstract
Temporal Action Detection (TAD), the task of localizing and classifying actions in untrimmed video, remains challenging due to action overlaps and variable action durations. Recent findings suggest that TAD performance is dependent on the structural design of transformers rather than on the self-attention mechanism. Building on this insight, we propose a refined feature extraction process through lightweight, yet effective operations. First, we employ a local branch that employs parallel convolutions with varying window sizes to capture both fine-grained and coarse-grained temporal features. This branch incorporates a gating mechanism to select the most relevant features. Second, we introduce a context branch that uses boundary frames as key-value pairs to analyze their relationship with the central frame through cross-attention. The proposed method captures temporal dependencies and improves contextual understanding. Evaluations of the gating mechanism and context branch on challenging datasets (THUMOS14 and EPIC-KITCHEN 100) show a consistent improvement over the baseline and existing methods.
Title:
UniDet3D: Multi-dataset Indoor 3D Object Detection
Authors: Maksim Kolodiazhnyi, Anna Vorontsova, Matvey Skripkin, Danila Rukhovich, Anton Konushin
Subjects: Subjects:
Computer Vision and Pattern Recognition (cs.CV)
Abstract
Growing customer demand for smart solutions in robotics and augmented reality has attracted considerable attention to 3D object detection from point clouds. Yet, existing indoor datasets taken individually are too small and insufficiently diverse to train a powerful and general 3D object detection model. In the meantime, more general approaches utilizing foundation models are still inferior in quality to those based on supervised training for a specific task. In this work, we propose \ours{}, a simple yet effective 3D object detection model, which is trained on a mixture of indoor datasets and is capable of working in various indoor environments. By unifying different label spaces, \ours{} enables learning a strong representation across multiple datasets through a supervised joint training scheme. The proposed network architecture is built upon a vanilla transformer encoder, making it easy to run, customize and extend the prediction pipeline for practical use. Extensive experiments demonstrate that \ours{} obtains significant gains over existing 3D object detection methods in 6 indoor benchmarks: ScanNet (+1.1 mAP50), ARKitScenes (+19.4 mAP25), S3DIS (+9.1 mAP50), MultiScan (+9.3 mAP50), 3RScan (+3.2 mAP50), and ScanNet++ (+2.7 mAP50). Code is available at this https URL .
Title:
Hermes: Memory-Efficient Pipeline Inference for Large Models on Edge Devices
Abstract
The application of Transformer-based large models has achieved numerous success in recent years. However, the exponential growth in the parameters of large models introduces formidable memory challenge for edge deployment. Prior works to address this challenge mainly focus on optimizing the model structure and adopting memory swapping methods. However, the former reduces the inference accuracy, and the latter raises the inference latency. This paper introduces PIPELOAD, a novel memory-efficient pipeline execution mechanism. It reduces memory usage by incorporating dynamic memory management and minimizes inference latency by employing parallel model loading. Based on PIPELOAD mechanism, we present Hermes, a framework optimized for large model inference on edge devices. We evaluate Hermes on Transformer-based models of different sizes. Our experiments illustrate that Hermes achieves up to 4.24 X increase in inference speed and 86.7% lower memory consumption than the state-of-the-art pipeline mechanism for BERT and ViT models, 2.58 X increase in inference speed and 90.3% lower memory consumption for GPT-style models.
Title:
AttentionX: Exploiting Consensus Discrepancy In Attention from A Distributed Optimization Perspective
Abstract
In this paper, we extend the standard Attention in transformer by exploiting the consensus discrepancy from a distributed optimization perspective, referred to as AttentionX. It is noted that %the popular distributed optimization algorithm \cite{Boyd11ADMM} and the primal-dual method of multipliers (PDMM) \cite{Zhang16PDMM} is designed to iteratively solve a broad class of distributed optimization problems over a pear-to-pear (P2P) network, where neighbouring nodes gradually reach consensus as specified by predefined linear edge-constraints in the optimization process. In particular, at each iteration of PDMM, each node in a network first performs information-gathering from neighbours and then performs local information-fusion. From a high-level point of view, the $KQ$-softmax-based weighted summation of $V$-representations in Attention corresponds information-gathering from neighbours while the feature-processing via the feed-forward network (FFN) in transformer corresponds to local information fusion. PDMM exploits the Lagrangian multipliers to capture the historical consensus discrepancy in the form of residual errors of the linear edge-constraints, which plays a crucial role for the algorithm to converge. Inspired by PDMM, we propose AttentionX to incorporate the consensus discrepancy in the output update-expression of the standard Attention. The consensus discrepancy in AttentionX refers to the difference between the weighted summation of $V$-representations and scaled $V$-representions themselves. Experiments on ViT and nanoGPT show promising performance.
Title:
Enhancing Sequential Music Recommendation with Personalized Popularity Awareness
Authors: Davide Abbattista, Vito Walter Anelli, Tommaso Di Noia, Craig Macdonald, Aleksandr Vladimirovich Petrov
Abstract
In the realm of music recommendation, sequential recommender systems have shown promise in capturing the dynamic nature of music consumption. Nevertheless, traditional Transformer-based models, such as SASRec and BERT4Rec, while effective, encounter challenges due to the unique characteristics of music listening habits. In fact, existing models struggle to create a coherent listening experience due to rapidly evolving preferences. Moreover, music consumption is characterized by a prevalence of repeated listening, i.e., users frequently return to their favourite tracks, an important signal that could be framed as individual or personalized popularity. This paper addresses these challenges by introducing a novel approach that incorporates personalized popularity information into sequential recommendation. By combining user-item popularity scores with model-generated scores, our method effectively balances the exploration of new music with the satisfaction of user preferences. Experimental results demonstrate that a Personalized Most Popular recommender, a method solely based on user-specific popularity, outperforms existing state-of-the-art models. Furthermore, augmenting Transformer-based models with personalized popularity awareness yields superior performance, showing improvements ranging from 25.2% to 69.8%. The code for this paper is available at this https URL.
Title:
Theory, Analysis, and Best Practices for Sigmoid Self-Attention
Authors: Jason Ramapuram, Federico Danieli, Eeshan Dhekane, Floris Weers, Dan Busbridge, Pierre Ablin, Tatiana Likhomanenko, Jagrit Digani, Zijin Gu, Amitis Shidani, Russ Webb
Abstract
Attention is a key part of the transformer architecture. It is a sequence-to-sequence mapping that transforms each sequence element into a weighted sum of values. The weights are typically obtained as the softmax of dot products between keys and queries. Recent work has explored alternatives to softmax attention in transformers, such as ReLU and sigmoid activations. In this work, we revisit sigmoid attention and conduct an in-depth theoretical and empirical analysis. Theoretically, we prove that transformers with sigmoid attention are universal function approximators and benefit from improved regularity compared to softmax attention. Through detailed empirical analysis, we identify stabilization of large initial attention norms during the early stages of training as a crucial factor for the successful training of models with sigmoid attention, outperforming prior attempts. We also introduce FLASHSIGMOID, a hardware-aware and memory-efficient implementation of sigmoid attention yielding a 17% inference kernel speed-up over FLASHATTENTION2 on H100 GPUs. Experiments across language, vision, and speech show that properly normalized sigmoid attention matches the strong performance of softmax attention on a wide range of domains and scales, which previous attempts at sigmoid attention were unable to fully achieve. Our work unifies prior art and establishes best practices for sigmoid attention as a drop-in softmax replacement in transformers.
Title:
Accelerating Training with Neuron Interaction and Nowcasting Networks
Authors: Boris Knyazev, Abhinav Moudgil, Guillaume Lajoie, Eugene Belilovsky, Simon Lacoste-Julien
Abstract
Neural network training can be accelerated when a learnable update rule is used in lieu of classic adaptive optimizers (e.g. Adam). However, learnable update rules can be costly and unstable to train and use. A simpler recently proposed approach to accelerate training is to use Adam for most of the optimization steps and periodically, only every few steps, nowcast (predict future) parameters. We improve this approach by Neuron interaction and Nowcasting (NiNo) networks. NiNo leverages neuron connectivity and graph neural networks to more accurately nowcast parameters by learning in a supervised way from a set of training trajectories over multiple tasks. We show that in some networks, such as Transformers, neuron connectivity is non-trivial. By accurately modeling neuron connectivity, we allow NiNo to accelerate Adam training by up to 50\% in vision and language tasks.
Keyword: scene understanding
Title:
RCNet: Deep Recurrent Collaborative Network for Multi-View Low-Light Image Enhancement
Abstract
Scene observation from multiple perspectives would bring a more comprehensive visual experience. However, in the context of acquiring multiple views in the dark, the highly correlated views are seriously alienated, making it challenging to improve scene understanding with auxiliary views. Recent single image-based enhancement methods may not be able to provide consistently desirable restoration performance for all views due to the ignorance of potential feature correspondence among different views. To alleviate this issue, we make the first attempt to investigate multi-view low-light image enhancement. First, we construct a new dataset called Multi-View Low-light Triplets (MVLT), including 1,860 pairs of triple images with large illumination ranges and wide noise distribution. Each triplet is equipped with three different viewpoints towards the same scene. Second, we propose a deep multi-view enhancement framework based on the Recurrent Collaborative Network (RCNet). Specifically, in order to benefit from similar texture correspondence across different views, we design the recurrent feature enhancement, alignment and fusion (ReEAF) module, in which intra-view feature enhancement (Intra-view EN) followed by inter-view feature alignment and fusion (Inter-view AF) is performed to model the intra-view and inter-view feature propagation sequentially via multi-view collaboration. In addition, two different modules from enhancement to alignment (E2A) and from alignment to enhancement (A2E) are developed to enable the interactions between Intra-view EN and Inter-view AF, which explicitly utilize attentive feature weighting and sampling for enhancement and alignment, respectively. Experimental results demonstrate that our RCNet significantly outperforms other state-of-the-art methods. All of our dataset, code, and model will be available at this https URL.
Title:
Future Does Matter: Boosting 3D Object Detection with Temporal Motion Estimation in Point Cloud Sequences
Authors: Rui Yu, Runkai Zhao, Cong Nie, Heng Wang, HuaiCheng Yan, Meng Wang
Subjects: Subjects:
Computer Vision and Pattern Recognition (cs.CV)
Abstract
Accurate and robust LiDAR 3D object detection is essential for comprehensive scene understanding in autonomous driving. Despite its importance, LiDAR detection performance is limited by inherent constraints of point cloud data, particularly under conditions of extended distances and occlusions. Recently, temporal aggregation has been proven to significantly enhance detection accuracy by fusing multi-frame viewpoint information and enriching the spatial representation of objects. In this work, we introduce a novel LiDAR 3D object detection framework, namely LiSTM, to facilitate spatial-temporal feature learning with cross-frame motion forecasting information. We aim to improve the spatial-temporal interpretation capabilities of the LiDAR detector by incorporating a dynamic prior, generated from a non-learnable motion estimation model. Specifically, Motion-Guided Feature Aggregation (MGFA) is proposed to utilize the object trajectory from previous and future motion states to model spatial-temporal correlations into gaussian heatmap over a driving sequence. This motion-based heatmap then guides the temporal feature fusion, enriching the proposed object features. Moreover, we design a Dual Correlation Weighting Module (DCWM) that effectively facilitates the interaction between past and prospective frames through scene- and channel-wise feature abstraction. In the end, a cascade cross-attention-based decoder is employed to refine the 3D prediction. We have conducted experiments on the Waymo and nuScenes datasets to demonstrate that the proposed framework achieves superior 3D detection performance with effective spatial-temporal feature learning.
Keyword: human object interaction
There is no result
Keyword: visual relation detection
There is no result
Keyword: object detection
Title:
Title:
Title:
Title:
Title:
Keyword: transformer
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Keyword: scene understanding
Title:
Title:
Keyword: visual reasoning
There is no result