By Duy-Kien Nguyen, Jihong Ju, Olaf Booij, Martin R. Oswald, Cees Snoek.
This repository is an official implementation of the paper BoxeR: Box-Attention for 2D and 3D Transformers.
TL; DR. BoxeR is a Transformer-based network for end-to-end 2D object detection and instance segmentation, along with 3D object detection. The core of the network is Box-Attention which predicts regions of interest to attend by learning the transformation (translation, scaling, and rotation) from reference windows, yielding competitive performance on several vision tasks.
Abstract. In this paper, we propose a simple attention mechanism, we call box-attention. It enables spatial interaction between grid features, as sampled from boxes of interest, and improves the learning capability of transformers for several vision tasks. Specifically, we present BoxeR, short for Box Transformer, which attends to a set of boxes by predicting their transformation from a reference window on an input feature map. The BoxeR computes attention weights on these boxes by considering its grid structure. Notably, BoxeR-2D naturally reasons about box information within its attention module, making it suitable for end-to-end instance detection and segmentation tasks. By learning invariance to rotation in the box-attention module, BoxeR-3D is capable of generating discriminative information from a bird's-eye view plane for 3D end-to-end object detection. Our experiments demonstrate that the proposed BoxeR-2D achieves state-of-the-art results on COCO detection and instance segmentation. Besides, BoxeR-3D improves over the end-to-end 3D object detection baseline and already obtains a compelling performance for the vehicle category of Waymo Open, without any class-specific optimization.
This project is released under the MIT License.
If you find BoxeR useful in your research, please consider citing:
@article{nguyen2021boxer,
title={BoxeR: Box-Attention for 2D and 3D Transformers},
author={Duy{-}Kien Nguyen and Jihong Ju and Olaf Booij and Martin R. Oswald and Cees G. M. Snoek},
journal={arXiv preprint arXiv:2111.13087},
year={2021}
}
Name | param (M) |
infer time (fps) |
box AP |
box AP-S |
box AP-M |
box AP-L |
segm AP |
segm AP-S |
segm AP-M |
segm AP-L |
---|---|---|---|---|---|---|---|---|---|---|
BoxeR-R50-3x | 40.1 | 12.5 | 50.3 | 33.4 | 53.3 | 64.4 | 42.9 | 22.8 | 46.1 | 61.7 |
BoxeR-R101-3x | 59.0 | 10.0 | 50.7 | 33.4 | 53.8 | 65.7 | 43.3 | 23.5 | 46.4 | 62.5 |
BoxeR-R101-5x | 59.0 | 10.0 | 51.9 | 34.2 | 55.8 | 67.1 | 44.3 | 24.7 | 48.0 | 63.8 |
-
Linux, CUDA>=11, GCC>=5.4
-
Python>=3.8
We recommend you to use Anaconda to create a conda environment:
conda create -n boxer python=3.8
Then, activate the environment:
conda activate boxer
-
PyTorch>=1.10.1, torchvision>=0.11.2 (following instructions here)
For example, you could install pytorch and torchvision as following:
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
-
Other requirements & Compilation
python -m pip install -e BoxeR
You can test the CUDA operators (box and instance attention) by running
python tests/box_attn_test.py python tests/instance_attn_test.py
The datasets are assumed to exist in a directory specified by the environment variable $E2E_DATASETS.
If the environment variable is not specified, it will be set to be .data
.
Under this directory, detectron2 will look for datasets in the structure described below.
$E2E_DATASETS/
├── coco/
└── waymo/
For COCO Detection and Instance Segmentation, please download COCO 2017 dataset and organize them as following:
$E2E_DATASETS/
└── coco/
├── annotation/
├── instances_train2017.json
├── instances_val2017.json
└── image_info_test-dev2017.json
├── image/
├── train2017/
├── val2017/
└── test2017/
└── vocabs/
└── coco_categories.txt - the mapping from coco categories to indices.
The coco_categories.txt
can be downloaded here.
For Waymo Detection, please download Waymo Open dataset and organize them as following:
$E2E_DATASETS/
└── waymo/
├── infos/
├── dbinfos_train_1sweeps_withvelo.pkl
├── infos_train_01sweeps_filter_zero_gt.pkl
└── infos_val_01sweeps_filter_zero_gt.pkl
└── lidars/
├── gt_database_1sweeps_withvelo/
├── CYCLIST/
├── VEHICLE/
└── PEDESTRIAN/
├── train/
├── annos/
└── lidars/
└── val/
├── annos/
└── lidars/
You can generate data files for our training and evaluation from raw data by running create_gt_database.py
and create_imdb
in tools/preprocess
.
Our script is able to automatically detect the number of available gpus on a single node. It works best with Slurm system when it can auto-detect the number of available gpus along with nodes. The command for training BoxeR is simple as following:
python tools/run.py --config ${CONFIG_PATH} --model ${MODEL_TYPE} --task ${TASK_TYPE}
For example,
- COCO Detection
python tools/run.py --config e2e/config/COCO-Detection/boxer2d_R_50_3x.yaml --model boxer2d --task detection
- COCO Instance Segmentation
python tools/run.py --config e2e/config/COCO-InstanceSegmentation/boxer2d_R_50_3x.yaml --model boxer2d --task detection
- Waymo Detection,
python tools/run.py --config e2e/config/Waymo-Detection/boxer3d_pointpillar.yaml --model boxer3d --task detection3d
- If your file system is slow to read images but your memory is huge, you may consider enabling 'cache_mode' option to load whole dataset into memory at the beginning of training:
python tools/run.py --config ${CONFIG_PATH} --model ${MODEL_TYPE} --task ${TASK_TYPE} dataset_config.${TASK_TYPE}.cache_mode=True
- If your GPU memory does not fit the batch size, you may consider to use 'iter_per_update' to perform gradient accumulation:
python tools/run.py --config ${CONFIG_PATH} --model ${MODEL_TYPE} --task ${TASK_TYPE} training.iter_per_update=2
- Our code also supports mixed precision training. It is recommended to use when you GPUs architecture can perform fast FP16 operations:
python tools/run.py --config ${CONFIG_PATH} --model ${MODEL_TYPE} --task ${TASK_TYPE} training.use_fp16=(float16 or bfloat16)
You can get the config file and pretrained model of BoxeR, then run following command to evaluate it on COCO 2017 validation/test set:
python tools/run.py --config ${CONFIG_PATH} --model ${MODEL_TYPE} --task ${TASK_TYPE} training.run_type=(val or test or val_test)
For Waymo evaluation, you need to additionally run the script e2edet/evaluate/waymo_eval.py
from the root folder to get the final result.
You can get the statistics of BoxeR (fps, flops, # parameters) by running tools/analyze.py
from the root folder.
python tools/analyze.py --config-path save/COCO-InstanceSegmentation/boxer2d_R_101_3x.yaml --model-path save/COCO-InstanceSegmentation/boxer2d_final.pth --tasks speed flop parameter
The notebook for BoxeR-2D visualization is provided in tools/visualization/BoxeR_2d_segmentation.ipynb
.