[YOLOv4] Discussion

[ElectronicElephant]

Hi @hellock and @xvjiarui ,

As the YOLOv3 implementation is going to an end (finally 😄), @WenqiangX and I are planning to work on YOLOv4.

However, compared with YOLOv3, YOLOv4 is notorious for its bags of tricks, many of which are not implemented in mm-detection yet. So, we open this issue in the hope of discussing some implementation details with mm-lab.

First, let us summarize the tricks used in YOLOv4.

Already in mmdetection

1. CIoU-Loss https://github.com/open-mmlab/mmdetection/pull/3151

2. Label Smoothing Included in YOLOv3

3. CosineAnnealingScheduler https://github.com/open-mmlab/mmdetection/issues/3227

4. Random training shapes Included in Dataloader

Waiting for merge

1. Mosaic Data Aug https://github.com/open-mmlab/mmdetection/pull/3389 Note that augmentation for mask and segmentation and blur augmentation are not implemented.

2. Efficient-Det YOLOv4 may rely on the neck of EfficientDet.

Not implemented yet

• [ ] CutMix

• [ ] DropBlock regularization

• [ ] Mish activation

• [ ] CSP

• [ ] Multi-input weighted residual connections

• [ ] CmBN

• [ ] Self Adversarial Training

• [ ] Eliminate grid sensitivity

• [x] Multiple anchors for a single gt Just change the code in YOLOv3Head. Should not be a problem.

• [ ] Optimal Hyper-parameters

• [x] SPP Should not be a problem.

• [ ] SAM-block

• [ ] PAN path-aggregation block

• [ ] DIoU-NMS

Our concern

1. Dependency.

If there is an ETA of EfficientDet and https://github.com/open-mmlab/mmdetection/pull/3389 ?

2. C++ or PyTorch for some tricks, including DIoU-NMS.

Should it be adopted in C++ code in mmcv, or we can use PyTorch code instead (which might be slower)?

3. Customized activation functions.

If we stick to use ConvModule of mmcv, how to use Mish-activation in an elegant manner?

---

Since the implementation of YOLOv4 is still at an early stage, the concern might grow, and we are glad to hear the voice from mm-lab. Cheers.

[hellock]

• EfficientDet still have a long way to go since the performance has not been reproduced. Mosaic augmentation should be merged in one (or two) week.
• I prefer implementing DIoU-NMS with c++ in mmcv, but if the speed difference is not much, python codes are also acceptable.
• It is simple to implement a new activation function in mmcv, like HSwish. ConvModule supports new activation methods without modifying the source code.

[hiyyg]

Hi @ElectronicElephant, I am also interested in having yolov4 in mmdetection. How is it going now?

[WenqiangX]

Hi @hiyyg we will pull a request soon (maybe at the end of Sep.), most of the functionality has been implemented.

[GreenTeaHua]

great, mark.

[WenqiangX]

Due to the computational resources limitation, we are not going to PR the YOLOv4 as planned. But we notice that an official implemention is merged in https://github.com/open-mmlab/mmdetection/pull/3824. @GreenTeaHua