This repository is based on VoVNet-v2

Faster R-CNN on SKU-110K dataset

Note

We measure the inference time of all models with batch size 1 on the same RTX2080Ti GPU machine.

• pytorch1.4.0

• CUDA 10.2

• cuDNN 7.3

Lightweight with FPNLite

• 64 FPN.OUT_CHANNELS = 64

• DW and Slim denote depthwise separable convolution and a thiner model with half the channel size, respectively.

FPN

Using this command with --num-gpus 1

python /path/to/sku110/train_net.py --config-file /path/to/sku110/configs/<config.yaml> --eval-only --num-gpus 1 MODEL.WEIGHTS <model.pth>

Installation

As this repository is implemented as a extension form (detectron2/projects) upon detectron2, you just install detectron2 following INSTALL.md.

Prepare for SKU-110K dataset:

• To download dataset, please visit here

• Extract the file downloaded to datasets/sku110/images

• Extract datasets/sku110/Annotations.zip, there are 2 folders Annotations and ImageSets

Training

To train a model, run

python /path/to/sku110/train_net.py --config-file /path/to/sku110/configs/<config.yaml>

For example, to launch end-to-end Faster R-CNN training with VoVNetV2-39 backbone on 8 GPUs, one should execute:

python /path/to/sku110/train_net.py --config-file /path/to/sku110/configs/faster_rcnn_V_39_FPN_3x.yaml --num-gpus 8

Evaluation

Model evaluation can be done similarly:

python /path/to/sku110/train_net.py --config-file /path/to/sku110/configs/faster_rcnn_V_39_FPN_3x.yaml --eval-only MODEL.WEIGHTS <model.pth>

Visualization

To visual the result, run

python /path/to/sku110/demo.py --config-file /path/to/sku110/configs/faster_rcnn_V_39_FPN_3x.yaml --input image.jpg --output image.jpg MODEL.WEIGHTS <model.pth>

Citing VoVNet

If you use VoVNet, please use the following BibTeX entry.

@inproceedings{lee2019energy,  title = {An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection},  author = {Lee, Youngwan and Hwang, Joong-won and Lee, Sangrok and Bae, Yuseok and Park, Jongyoul},  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},  year = {2019}}