Iteration Plan - Apr 2021

[dreamerlin]

Since MMAction2 v0.13.0 has been released, now it's time to plan for the next iteration, i.e. v0.14.0.
This will be a one-month iteration and our focus in this iteration is to support more algorithms and refactor some codes.

Algorithms:

• PoseC3D ([Feature] PoseC3D #786)
• TRN ([Feature] Support TRN #755)
• TDN

Dataset:

• AVA-Kinetics
• Charades ([Feature] Support Charades Dataset #788)

Misc:

- [ ] Entrypoint Refactoring
- [ ] Unify parameters initilization (#794) blocked by upstream

• Model Registry ([Refactoring] Unified parameters initialization mmcv#780)
• Registry conflict resolution ([Refactor] Use MMCV Model Registry #843)
• Video Transformer researching
• Support auto_fp16 using torch.cuda.amp

[irvingzhang0512]

• Test with ONNX model or TensorRT engine. [Improvement] Test with onnx models and TensorRT engines. #758
• Recognizers Comparision Comparisons for Recognizers #773
• Webcam Demo for Spatio-Temporal Action Detection models. [Feature] Support Webcam Demo for Spatio-temporal Action Detection Models #795
• [ ] YOWO for ava Working on deployement tools

[Deep-learning999]

I hope to provide pre-training of different frame lengths in slowonly and slowfast of ava-Kinetics, as well as pre-training of TimeSformer in ava2.1 and ava-Kinetics. I think the real-time performance and map should be improved.

由于MMAction2 v0.13.0已发布，现在是时候计划下一个迭代，即v0.14.0。
这将是一个月的迭代，我们在此迭代中的重点是支持更多算法并重构某些代码。

算法：

• 姿势C3D（＃786）
• TRN（＃755）
• TDN

数据集：

• AVA运动学
• 夏拉德（＃788）

其他：

OpenMMLab重构项目

• 入口点重构
• 模型注册表（open-mmlab / mmcv＃780）
• 统一参数初始化（＃794）
• 解决注册表冲突
• 视频变压器研究
• 使用torch.cuda.amp支持auto_fp16

I hope to provide pre-training of different frame lengths in slowonly and slowfast of ava-Kinetics, as well as pre-training of TimeSformer in ava2.1 and ava-Kinetics. I think the real-time performance and map should be improved.