search

vita-epfl / detection-attributes-fields

PyTorch implementation of "Detecting 32 Pedestrian Attributes for Autonomous Vehicles"
https://arxiv.org/abs/2012.02647
Other
31 stars 6 forks source link
attribute-recognition autonomous-vehicles computer-vision deep-learning jaad machine-learning multi-task-learning object-detection pedestrian-behavior road-crossing

readme

Object Detection and Attribute Recognition with Fields

PyTorch implementation of paper Detecting 32 Pedestrian Attributes for Autonomous Vehicles by Taylor Mordan (EPFL/VITA), Matthieu Cord (Sorbonne Université, valeo.ai), Patrick Pérez (valeo.ai) and Alexandre Alahi (EPFL/VITA).

Abstract

Detecting 32 Pedestrian Attributes for Autonomous Vehicles

Pedestrians are arguably one of the most safety-critical road users to consider for autonomous vehicles in urban areas. In this paper, we address the problem of jointly detecting pedestrians and recognizing 32 pedestrian attributes from a single image. These encompass visual appearance and behavior, and also include the forecasting of road crossing, which is a main safety concern. For this, we introduce a Multi-Task Learning (MTL) model relying on a composite field framework, which achieves both goals in an efficient way. Each field spatially locates pedestrian instances and aggregates attribute predictions over them. This formulation naturally leverages spatial context, making it well suited to low resolution scenarios such as autonomous driving. By increasing the number of attributes jointly learned, we highlight an issue related to the scales of gradients, which arises in MTL with numerous tasks. We solve it by normalizing the gradients coming from different objective functions when they join at the fork in the network architecture during the backward pass, referred to as fork-normalization. Experimental validation is performed on JAAD, a dataset providing numerous attributes for pedestrian analysis from autonomous vehicles, and shows competitive detection and attribute recognition results, as well as a more stable MTL training.

detection_schema

The model MTL-Fields learns multiple fields for both object detection and attribute recognition in a Multi-Task Learning way. Learning is done on full images with dedicated field and image-wise loss function for each task, and predictions are obtained at inference through a post-processing instance-wise decoding step that yields a bounding box and all attributes for each detected instance. This model is applied on dataset JAAD to detect up to 32 pedestrian attributes in an autonomous vehicle scenario.

The model MTL-Fields also contains a normalization of gradients during backward to solve gradient scale issues when learning numerous tasks.

Table of Contents

Installation

Clone this repository in order to use it.

# To clone the repository using HTTPS
git clone https://github.com/vita-epfl/detection-attributes-fields
cd detection-attributes-fields/

All dependencies can be found in the requirements.txt file.

# To install dependencies
pip3 install -r requirements.txt

This project has been tested with Python 3.7.7, PyTorch 1.9.1, CUDA 10.2 and OpenPifPaf 0.13.0.

Dataset

This project uses dataset JAAD for training and evaluation.

Please refer to JAAD documentation to download the dataset.

Interfaces

This project is implemented as an OpenPifPaf plugin module. As such, it benefits from all the core capabilities offered by OpenPifPaf, and only implements the additional functions it needs.

All the commands can be run through OpenPifPaf's interface using subparsers. Help can be obtained for any of them with option --help. More information can be found in OpenPifPaf documentation.

Training

Training is done using subparser openpifpaf.train.

Training on JAAD with all attributes can be run with the command:

python3 -m openpifpaf.train \
  --output <path/to/model.pt> \
  --dataset jaad \
  --jaad-root-dir <path/to/jaad/folder/> \
  --jaad-subset default \
  --jaad-training-set train \
  --jaad-validation-set val \
  --log-interval 10 \
  --val-interval 1 \
  --epochs 5 \
  --batch-size 4 \
  --lr 0.0005 \
  --lr-warm-up-start-epoch -1 \
  --weight-decay 5e-4 \
  --momentum 0.95 \
  --basenet fn-resnet50 \
  --pifpaf-pretraining \
  --detection-bias-prior 0.01 \
  --jaad-head-upsample 2 \
  --jaad-pedestrian-attributes all \
  --fork-normalization-operation power \
  --fork-normalization-duplicates 35 \
  --lambdas 7.0 7.0 7.0 7.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 \
  --attribute-regression-loss l1 \
  --attribute-focal-gamma 2 \
  --auto-tune-mtl

Arguments should be modified appropriately if needed.

More information about the options can be obtained with the command:

python3 -m openpifpaf.train --help

Evaluation

Evaluation of a checkpoint is done using subparser openpifpaf.eval.

Evaluation on JAAD with all attributes can be run with the command:

python3 -m openpifpaf.eval \
  --output <path/to/outputs> \
  --dataset jaad \
  --jaad-root-dir <path/to/jaad/folder/> \
  --jaad-subset default \
  --jaad-testing-set test \
  --checkpoint <path/to/checkpoint.pt> \
  --batch-size 1 \
  --jaad-head-upsample 2 \
  --jaad-pedestrian-attributes all \
  --head-consolidation filter_and_extend \
  --decoder instancedecoder:0 \
  --decoder-s-threshold 0.2 \
  --decoder-optics-min-cluster-size 10 \
  --decoder-optics-epsilon 5.0 \
  --decoder-optics-cluster-threshold 0.5

Arguments should be modified appropriately if needed.

Using option --write-predictions, a json file with predictions can be written as an additional output.

Using option --show-final-image, images with predictions displayed on them can be written in the folder given by option --save-all <path/to/image/folder/>. To also display ground truth annotations, add option --show-final-ground-truth.

More information about the options can be obtained with the command:

python3 -m openpifpaf.eval --help

Project structure

The code is organized as follows:

openpifpaf_detection_attributes/
├── datasets/
│   ├── jaad/
│   ├── (+ common files for datasets)
│   └── (add new datasets here)
└── models/
    ├── mtlfields/
    ├── (+ common files for models)
    └── (add new models here)

License

This project is built upon OpenPifPaf and shares the AGPL Licence.

This software is also available for commercial licensing via the EPFL Technology Transfer Office (https://tto.epfl.ch/, info.tto@epfl.ch).

Citation

If you use this project in your research, please cite the corresponding paper:

@article{mordan2021detecting,
  title={Detecting 32 Pedestrian Attributes for Autonomous Vehicles},
  author={Mordan, Taylor and Cord, Matthieu and P{\'e}rez, Patrick and Alahi, Alexandre},
  journal={IEEE Transactions on Intelligent Transportation Systems (T-ITS)},
  year={2021},
  doi={10.1109/TITS.2021.3107587}
}

Acknowledgements

We would like to thank Valeo for funding our work, and Sven Kreiss for the OpenPifPaf Plugin architecture.