Relation DETR
By Xiuquan Hou, Meiqin Liu, Senlin Zhang, Ping Wei, Badong Chen, Xuguang Lan.
This repo is the official implementation of Relation DETR: Exploring Explicit Position Relation Prior for Object Detection accepted to ECCV2024 (score 5444, oral presentation). [Arxiv paper link], [论文介绍], [代码讲解]
💖 If our Relation-DETR or SA-Det-100k dataset is helpful to your researches or projects, please star this repository. Thanks! 🤗
TODO
...Want more features? Open a Feature Request.
- [x] Support data augmentations from
albumentations. - [ ] Support Mosaic and Mixup data augmentation.
- [ ] More detailed docs for the code.
- [ ] Add a instruction about introducing our relation to other models.
- [ ] Support GradCam and feature visualization.
- [x] Upload more pretrained weights and training logs.
- [x] Update visualization code for MC.
- [x] Update Model ZOO.
Update
- [2024-09-08] Relation-DETR (Focal-Large) checkpoint pretrained on Object365 is now available here.
- [2024-08-15] We release config and checkpoint of DINO++ (DINO enhanced by our position relation).
- [2024-08-12] Relation-DETR is selected for Oral presentation in ECCV2024!
- [2024-08-11] The pretrained weight for Relation-DETR on SA-Det-100k are available here!
- [2024-08-07] Relation-DETR with FocalNet-large achieves
63.5APon COCO test-dev2017 after fine-tuned for 4 epochs on Object365, config and checkpoint are available now! - [2024-07-24] Upload SA-Det-100k dataset, see it in Hugging Face and Ai Studio.
- [2024-07-18] Upload Relation-DETR training logs for pretrained weights.
- [2024-07-18] We release the code for Relation-DETR, Relation-DETR with Swin-L achieves 58.1 AP!
- [2024-03-26] Code for Salience-DETR is available here.
- [2024-07-17] We release the checkpoint for Relation-DETR with ResNet-50 and Swin-L backbones, see Releases v1.0.0.
- [2024-07-01] Relation-DETR is accepted to ECCV2024. Welcome to your attention!
SA-Det-100k
SA-Det-100k is a large-scale class-agnostic object detection dataset for Research Purposes only. The dataset is based on a subset of SA-1B (see LICENSE), and all objects belong to the same category objects. Because it contains a large number of scenarios but does not provide class-specific annotations, we believe it may be a good choice to pre-training models for a variety of downstream tasks with different categories. The dataset contains about 100k images, and each image is resized using opencv-python so that the larger one of their height and width is 1333, which is consistent with the data augmentation commonly used to train COCO. The dataset can be found in:
Model ZOO
COCO
| Model | Backbone | Epoch | Download | mAP | AP50 | AP75 | APS | APM | APL |
|---|---|---|---|---|---|---|---|---|---|
| Relation DETR | ResNet50 | 12 | config / checkpoint / log | 51.7 | 69.1 | 56.3 | 36.1 | 55.6 | 66.1 |
| Relation DETR | Swin-L(IN-22K) | 12 | config / checkpoint | 57.8 | 76.1 | 62.9 | 41.2 | 62.1 | 74.4 |
| Relation DETR | ResNet50 | 24 | config / checkpoint / log | 52.1 | 69.7 | 56.6 | 36.1 | 56.0 | 66.5 |
| Relation DETR | Swin-L(IN-22K) | 24 | config / checkpoint / log | 58.1 | 76.4 | 63.5 | 41.8 | 63.0 | 73.5 |
| Relation-DETR† | Focal-L(IN-22K) | 4+24 | config / o365_checkpoint / checkpoint | 63.5 | 80.8 | 69.1 | 47.2 | 66.9 | 77.0 |
† means finetuned model on COCO after pretraining on Object365.
[Other DETR variants:] We integrate our position relation into existing DETR variants and generate enhanced versions of them. Note some of these weights are newly trained and may produce slightly different results from those reported in our paper. We mark these variants with ++ in the name to distinguish them from their original versions.
| Model | Backbone | Epoch | Download | mAP | AP50 | AP75 | APS | APM | APL |
|---|---|---|---|---|---|---|---|---|---|
| DINO++ | ResNet50 | 12 | config / checkpoint | 50.1 | 67.8 | 54.9 | 33.3 | 53.9 | 63.5 |
SA-Det-100k
| Model | Backbone | Epoch | Download | mAP | AP50 | AP75 | APS | APM | APL |
|---|---|---|---|---|---|---|---|---|---|
| DINO with VFL | ResNet50 | 12 | —— | 43.7 | 52.0 | 47.7 | 5.8 | 43.0 | 61.5 |
| Relation DETR | ResNet50 | 12 | config / checkpoint | 45.0 | 53.1 | 48.9 | 6.0 | 44.4 | 62.9 |
Get started
1. Installation
**We use the environment same as [Salience-DETR](https://arxiv.org/abs/2403.16131). You can skip the step if you have run Salience-DETR.** 1. Clone the repository: ```shell git clone https://github.com/xiuqhou/Relation-DETR cd Relation-DETR ``` 2. Install Pytorch and torchvision: ```shell conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch ``` 3. Install other requirements: ```shell pip install -r requirements.txt ```2. Prepare datasets
Download [COCO2017](https://cocodataset.org/) (and [SA-Det-100k](https://huggingface.co/datasets/xiuqhou/SA-Det-100k) optionally), put them in `data/` following the structure: ```shell data/ ├─coco/ │ ├── train2017/ │ ├── val2017/ │ └── annotations/ │ ├── instances_train2017.json │ └── instances_val2017.json │ └─sa_det_100k/ ├── train2017/ ├── val2017/ └── annotations/ ```3. Evaluate pretrained models
To evaluate a model with one or more GPUs, specify `CUDA_VISIBLE_DEVICES`, `dataset`, `model` and `checkpoint`. ```shell CUDA_VISIBLE_DEVICES=4. Evaluate exported json results
To evaluate a json results, specify `dataset` and `result`. The evaluation only needs CPU so you don't need to specify `CUDA_VISIBLE_DEVICES`. ```shell accelerate launch test.py --coco-path /path/to/coco --result /path/to/result.json ``` - To visualize predictions, specify `--show-dir` with a folder name. You can change the visualization style through `--font-scale`, `--box-thick`, `--fill-alpha`, `--text-box-color`, `--text-font-color`, `--text-alpha` parameters.5. Train a model
Use `CUDA_VISIBLE_DEVICES` to specify GPU/GPUs and run the following script to start training. If not specified, the script will use all available GPUs on the node to train. Before start training, modify parameters in [configs/train_config.py](configs/train_config.py). ```shell CUDA_VISIBLE_DEVICES=0 accelerate launch main.py # train with 1 GPU CUDA_VISIBLE_DEVICES=0,1 accelerate launch main.py # train with 2 GPUs ```5. Benchmark a model
To test the inference speed, memory cost and parameters of a model, use tools/benchmark_model.py. ```shell python tools/benchmark_model.py --model-config configs/relation_detr/relation_detr_resnet50_800_1333.py ```6. Export an ONNX model
For advanced users who want to deploy our model, we provide a script to export an ONNX file. ```shell python tools/pytorch2onnx.py \ --model-config /path/to/model.py \ --checkpoint /path/to/checkpoint.pth \ --save-file /path/to/save.onnx \ --simplify \ # use onnxsim to simplify the exported onnx file --verify # verify the error between onnx model and pytorch model ``` For inference using the ONNX file, see ONNXDetector in [tools/pytorch2onnx.py](tools/pytorch2onnx.py)License
Relation-DETR is released under the Apache 2.0 license. Please see the LICENSE file for more information.
Bibtex
If you find our work helpful for your research, please consider citing:
@inproceedings{hou2024relation,
title={Relation DETR: Exploring Explicit Position Relation Prior for Object Detection},
author={Hou, Xiuquan and Liu, Meiqin and Zhang, Senlin and Wei, Ping and Chen, Badong and Lan, Xuguang},
booktitle={European conference on computer vision},
year={2024},
organization={Springer}
}
@InProceedings{Hou_2024_CVPR,
author = {Hou, Xiuquan and Liu, Meiqin and Zhang, Senlin and Wei, Ping and Chen, Badong},
title = {Salience DETR: Enhancing Detection Transformer with Hierarchical Salience Filtering Refinement},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2024},
pages = {17574-17583}
}