Source code and evaluation scripts for our Pattern Recognition 2023 paper:

YOGA: Deep Object Detection in the Wild with Lightweight Feature Learning and Multiscale Attention

Link to paper
arXiv

Abstract

We introduce YOGA, a deep learning based yet lightweight object detection model that can operate on low-end edge devices while still achieving competitive accuracy. The YOGA architecture consists of a two-phase feature learning pipeline with a cheap linear transformation, which learns feature maps using only half of the convolution filters required by conventional convolutional neural networks. In addition, it performs multi-scale feature fusion in its neck using an attention mechanism instead of the naive concatenation used by conventional detectors. YOGA is a flexible model that can be easily scaled up or down by several orders of magnitude to fit a broad range of hardware constraints. We evaluate YOGA on COCO-val and COCO-testdev datasets with over 10 state-of-the-art object detectors. The results show that YOGA strikes the best trade-off between model size and accuracy (up to 22% increase of AP and 23–34% reduction of parameters and FLOPs), making it an ideal choice for deployment in the wild on low-end edge devices. This is further affirmed by our hardware implementation and evaluation on NVIDIA Jetson Nano.

Citation:

@article{pattern2023yoga,
  title={{YOGA}: Deep Object Detection in the Wild with Lightweight Feature Learning and Multiscale Attention},
  author={Sunkara, Raja and Luo, Tie},
  journal={Pattern Recognition},
  volume={139},
  pages={109451},
  year={2023},
  publisher={Elsevier},
  doi= {10.1016/j.patcog.2023.109451},
}

YOGA Building Block:

Installation

# Download the code 
git clone https://github.com/LabSAINT/YOGA

# Create an environment
cd SPD-Conv
conda create -n YOGA python==3.7.4
conda activate YOGA
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia
pip3 install -r requirements.txt

YOGA

YOGA has been trained and evaluated using the COCO-2017 dataset. Below are the pretrained models, evaluation, and training scripts

Pre-trained models

The table below gives an overview of the results of our models

$$\textbf{Model}$$	$$\textbf{AP}$$	$$\textbf{AP}_\textbf{S}$$	$$\textbf{Params (M)}$$	$$\textbf{FLOPs (B)}$$
YOGA-n	32.3	15.2	1.9	4.9
YOGA-s	40.7	23.0	7.6	16.6
YOGA-m	45.2	28.0	16.3	34.6
YOGA-l	48.9	31.8	33.6	71.8

Evaluation

The script val.py can be used to evaluate the pre-trained models

  $ python val.py --weights 'pth_models/YOGA-n.pt' --img 640 --iou 0.65 --half --batch-size 1 --data data/coco.yaml
  $ python val.py --weights 'pth_models/YOGA-s.pt' --img 640 --iou 0.65 --half --batch-size 1 --data data/coco.yaml
  $ python val.py --weights 'pth_models/YOGA-m.pt' --img 640 --iou 0.65 --half --batch-size 1 --data data/coco.yaml   

Training

The script train.py is used to train YOGA models

# nano model
python3 train.py --data coco.yaml --cfg ./models/YOGA-n.yaml --hyp ./data/hyps/YOGA-n.yaml --weights '' --batch-size 128 --epochs 300 --sync-bn --project YOGA --name YOGA-n --label-smoothing 0.01

# small model
python3 train.py --data coco.yaml --cfg ./models/YOGA-s.yaml --hyp ./data/hyps/YOGA-s.yaml --weights '' --batch-size 128 --epochs 300 --sync-bn --project YOGA --name YOGA-s --linear-lr

# medium model
python3 train.py --data coco.yaml --cfg ./models/YOGA-m.yaml --hyp ./data/hyps/YOGA-m.yaml --weights '' --batch-size 64 --epochs 300 --sync-bn --project YOGA --name YOGA-m

# large model

python3 train.py --data coco.yaml --cfg ./models/YOGA-l.yaml --hyp ./data/hyps/YOGA-m.yaml --weights '' --batch-size 20 --epochs 250 --sync-bn --project YOGA --name YOGA-l

AttributeError Issue Resolution

If you encounter the error: AttributeError: 'Upsample' object has no attribute 'recompute_scale_factor' A solution has been provided in the YOLOv5 GitHub repository.

Details of the Issue:

• Error: AttributeError: 'Upsample' object has no attribute 'recompute_scale_factor'
• Resolution: Amend the line containing recompute_scale_factor=self.recompute_scale_factor in E:\condaaa\lib\site-packages\torch\nn\modules\upsampling.py file.

For a detailed discussion and context, please refer to this GitHub issue comment.