Darknet: RefineDetLite: A Lightweight One-stage Object Detection Framework for CPU-only Devices

[WongKinYiu]

RefineDetLite: A Lightweight One-stage Object Detection Framework for CPU-only Devices

• paper https://arxiv.org/abs/1911.08855
• Res2Net + LightHead + RefineDet

Plan to implement Res2NetLite72.

[WongKinYiu]

Start training Res2NetLite72...

[WongKinYiu]

| Model | BFLOPs | Top-1 | Top-5 | cfg | weight |
| :---- | :----: | :---: | :---: | :-: | :----: |
| Res2NetLite-72 | 5.19 | 74.7 | 92.1 | cfg | weight |
| CSPResNeXt-50-Elastic | 5.45 | 77.2 | 93.8 | cfg | weight |

[AlexeyAB]

@WongKinYiu

Did you compare speed/accuracy rather than BFlops/accuracy of both networks?

Both nets uses grouped-convolutional: Res2NetLite-72 (groups=2) and CSPResNeXt-50-Elastic (groups=16)

[WongKinYiu]

@AlexeyAB Hello,

I have no free gpu now, could you help for testing the inference time?

[WongKinYiu]

@AlexeyAB

I find that res2net, espnet, espnetv2, and centermask are very similar.

mayebe @glenn-jocher will be interested in espnetv2.
https://github.com/sacmehta/ESPNetv2-COREML
it can do real-time semantic segmentation on iPhone7.

maybe you will be interested in centermask.
it is a real-time anchor free instance segmentation which is better than YOLACT.

[uday60]

@WongKinYiu
Is RefineDetLite available to train ?

[WongKinYiu]

@uday60

i only implement the backbone of refinedetlite.
cfg and weights are here.

[uday60]

@WongKinYiu
I can train on MSCOCO if you can share the network configuration file for both Res2NetLite-72 based yolo and CSPResNeXt-50-Elastic based yolo.