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

Backbone	Param.	lr sched	inference time	AP	AP75	AP50	download
MobileNetV2-0.5-64	N/A	1x	0.033	43.31	44.66	78.08	model | metrics
MobileNetV2-0.5	N/A	1x	0.037	42.93	44.27	77.31	model | metrics
MobileNetV2	3.5M	3x	0.031	52.11	58.72	85.98	model | metrics
MobileNetV2	3.5M	1x	0.031	51.20	56.93	85.71	model | metrics
MobileNetV2-FLGC	N/A	1x	0.030	50.59	56.05	85.21	model | metrics
ShuffleNetV2-0.5	N/A	1x	0.039	48.24	52.95	82.10	model | metrics
ShuffleNetV2	N/A	1x	0.028	52.60	59.55	86.19	model | metrics
V2-19	11.2M	1x	0.034	41.46	44.97	71.32	model | metrics
V2-19-DW	6.5M	1x	N/A	N/A	N/A	N/A	model | metrics
V2-19-Slim	3.1M	1x	0.027	47.68	51.47	82.36	model | metrics
V2-19-Slim-DW	1.8M	3x	N/A	N/A	N/A	N/A	model | metrics

• 64 FPN.OUT_CHANNELS = 64
• DW and Slim denote depthwise separable convolution and a thiner model with half the channel size, respectively.

FPN

Backbone	Param.	lr sched	inference time	AP	AP75	AP50	download
V2-19-FPN	37.6M	3x	N/A	N/A	N/A	N/A	model | metrics
R-50-FPN	51.2M	3x	N/A	N/A	N/A	N/A	model | metrics
V2-39-FPN	52.6M	3x	0.071	51.47	57.5	85.5	model | metrics

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}
}