ViDT: An Efficient and Effective Fully Transformer-based Object Detector

by Hwanjun Song1, Deqing Sun2, Sanghyuk Chun1, Varun Jampani2, Dongyoon Han1,
Byeongho Heo1, Wonjae Kim1, and Ming-Hsuan Yang2,3

1 NAVER AI Lab, 2 Google Research, 3 University California Merced

ViDT: Vision and Detection Transformers

Highlight

ViDT is an end-to-end fully transformer-based object detector, which directly produces predictions without using convolutional layers. Our main contributions are summarized as follows:

  • ViDT introduces a modified attention mechanism, named Reconfigured Attention Module (RAM), that facilitates any ViT variant to handling the appened [DET] and [PATCH] tokens for a standalone object detection. Thus, we can modify the lastest Swin Transformer backbone with RAM to be an object detector and obtain high scalability using its local attetention mechanism with linear complexity.

  • ViDT adopts a lightweight encoder-free neck architecture to reduce the computational overhead while still enabling the additional optimization techniques on the neck module. As a result, ViDT obtains better performance than neck-free counterparts.

  • We introdcue a new concept of token matching for knowledge distillation, which brings additional performance gains from a large model to a small model without compromising detection efficiency.

Architectural Advantages. First, ViDT enables to combine Swin Transformer and the sequent-to-sequence paradigm for detection. Second, ViDT can use the multi-scale features and additional techniques without a significant computation overhead. Therefore, as a fully transformer-based object detector, ViDT facilitates better integration of vision and detection transformers.

Component Summary. There are four components: (1) RAM to extend Swin Transformer as a standalone object detector, (2) the neck decoder to exploit multi-scale features with two additional techniques, auxiliary decoding loss and iterative box refinement, (3) knowledge distillation to benefit from a large model, and (4) decoding layer drop to further accelerate inference speed.

Evaluation

Index: [A. ViT Backbone], [B. Main Results], [C. Complete Analysis]

|--- A. ViT Backbone used for ViDT
|--- B. Main Results in the ViDT Paper
     |--- B.1. ViDT for 50 and 150 Epochs
     |--- B.2. Distillation with Token Matching
|--- C. Complete Component Analysis

A. ViT Backbone used for ViDT

Backbone and Size Training Data Epochs Resulution Params ImageNet Acc. Checkpoint
Swin-nano ImageNet-1K 300 224 6M 74.9% Github
Swin-tiny ImageNet-1K 300 224 28M 81.2% Github
Swin-small ImageNet-1K 300 224 50M 83.2% Github
Swin-base ImageNet-22K 90 224 88M 86.3% Github

B. Main Results in the ViDT Paper

In main experiments, auxiliary decoding loss and iterative box refinement were used as the auxiliary techniques on the neck structure.
The efficiacy of distillation with token mathcing and decoding layer drop are verified independently in Compelete Component Analysis.
All the models were re-trained with the final version of source codes. Thus, the value may be very slightly different from those in the paper.

B.1. VIDT for 50 and 150 epochs
Backbone Epochs AP AP50 AP75 AP_S AP_M AP_L Params FPS Checkpoint / Log
Swin-nano 50 (150) 40.4 (42.6) 59.9 (62.2) 43.0 (45.7) 23.1 (24.9) 42.8 (45.4) 55.9 (59.1) 16M 20.0 Github / Log
(Github / Log)
Swin-tiny 50 (150) 44.9 (47.2) 64.7 (66.7) 48.3 (51.4) 27.5 (28.4) 47.9 (50.2) 61.9 (64.7) 38M 17.2 Github / Log
(Github / Log)
Swin-small 50 (150) 47.4 (48.8) 67.7 (68.8) 51.2 (53.0) 30.4 (30.7) 50.7 (52.0) 64.6 (65.9) 60M 12.1 Github / Log
(Github / Log)
Swin-base 50 (150) 49.4 (50.4) 69.6 (70.4) 53.4 (54.8) 31.6 (34.1) 52.4 (54.2) 66.8 (67.4) 0.1B 9.0 Github / Log
(Github / Log)
B.2. Distillation with Token Matching (Coefficient 4.0)

All the models are trained for 50 epochs with distillation.

Teacher ViDT (Swin-base) trained for 50 epochs
Student ViDT (Swin-nano) ViDT (Swin-tiny) ViDT (Swin-Small)
Coefficient = 0.0 40.4 44.9 47.4
Coefficient = 4.0 41.8 (Github / Log) 46.6 (Github / Log) 49.2 (Github / Log)

C. Complete Component Analysis

We combined the four proposed components (even with distillation with token matching and decoding layer drop) to achieve high accuracy and speed for object detection. For distillation, ViDT (Swin-base) trained for 50 epochs was used for all models.

Component Swin-nano Swin-tiny Swin-small
# RAM Neck Distil Drop AP Params FPS AP Params FPS AP Params FPS
(1) ✔️ 28.7 7M 36.5 36.3 29M 28.6 41.6 52M 16.8
(2) ✔️ ✔️ 40.4 16M 20.0 44.9 38M 17.2 47.4 60M 12.1
(3) ✔️ ✔️ ✔️ 41.8 16M 20.0 46.6 38M 17.2 49.2 60M 12.1
(4) ✔️ ✔️ ✔️ ✔️ 41.6 13M 23.0 46.4 35M 19.5 49.1 58M 13.0

Requirements

This codebase has been developed with the setting used in Deformable DETR:
Linux, CUDA>=9.2, GCC>=5.4, Python>=3.7, PyTorch>=1.5.1, and torchvision>=0.6.1.

We recommend you to use Anaconda to create a conda environment:

conda create -n deformable_detr python=3.7 pip
conda activate deformable_detr
conda install pytorch=1.5.1 torchvision=0.6.1 cudatoolkit=9.2 -c pytorch

Compiling CUDA operators for deformable attention

cd ./ops
sh ./make.sh
# unit test (should see all checking is True)
python test.py

Other requirements

pip install -r requirements.txt

Training

We used the below commands to train ViDT models with a single node having 8 NVIDIA V100 GPUs.

Run this command to train the ViDT (Swin-nano) model in the paper : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_nano \
       --epochs 50 \
       --lr 1e-4 \
       --min-lr 1e-7 \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --coco_path /path/to/coco \
       --output_dir /path/for/output
Run this command to train the ViDT (Swin-tiny) model in the paper : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_tiny \
       --epochs 50 \
       --lr 1e-4 \
       --min-lr 1e-7 \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --coco_path /path/to/coco \
       --output_dir /path/for/output
Run this command to train the ViDT (Swin-small) model in the paper : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_small \
       --epochs 50 \
       --lr 1e-4 \
       --min-lr 1e-7 \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --coco_path /path/to/coco \
       --output_dir /path/for/output
Run this command to train the ViDT (Swin-base) model in the paper : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_base_win7_22k \
       --epochs 50 \
       --lr 1e-4 \
       --min-lr 1e-7 \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --coco_path /path/to/coco \
       --output_dir /path/for/output

When a large pre-trained ViDT model is available, distillation with token matching can be applied for training a smaller ViDT model.

Run this command when training ViDT (Swin-nano) using a large ViDT (Swin-base) via Knowledge Distillation : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_nano \
       --epochs 50 \
       --lr 1e-4 \
       --min-lr 1e-7 \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --distil_model vidt_base \
       --distil_path /path/to/vidt_base (or url) \
       --coco_path /path/to/coco \
       --output_dir /path/for/output

Evaluation

Run this command to evaluate the ViDT (Swin-nano) model on COCO : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \ 
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_nano \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --coco_path /path/to/coco \
       --resume /path/to/vidt_nano \
       --pre_trained none \
       --eval True
Run this command to evaluate the ViDT (Swin-tiny) model on COCO : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_tiny \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --coco_path /path/to/coco \
       --resume /path/to/vidt_tiny\
       --pre_trained none \
       --eval True
Run this command to evaluate the ViDT (Swin-small) model on COCO : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_small \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --coco_path /path/to/coco \
       --resume /path/to/vidt_small \
       --pre_trained none \
       --eval True
Run this command to evaluate the ViDT (Swin-base) model on COCO : 


python -m torch.distributed.launch \
       --nproc_per_node=8 \
       --nnodes=1 \
       --use_env main.py \
       --method vidt \
       --backbone_name swin_base_win7_22k \
       --batch_size 2 \
       --num_workers 2 \
       --aux_loss True \
       --with_box_refine True \
       --coco_path /path/to/coco \
       --resume /path/to/vidt_base \
       --pre_trained none \
       --eval True

Citation

Please consider citation if our paper is useful in your research.

@article{song2021vidt,
  title={ViDT: An Efficient and Effective Fully Transformer-based Object Detector},
  author={Song, Hwanjun and Sun, Deqing and Chun, Sanghyuk and Jampani, Varun and Han, Dongyoon and Heo, Byeongho and Kim, Wonjae and Yang, Ming-Hsuan},
  journal={arXiv preprint arXiv:2110.03921},
  year={2021}
}

License

Copyright 2021-present NAVER Corp.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

GitHub

View Github