open-mmlab
动作分类不准确，也比较奇怪
Branch
main branch (1.x version, such as v1.0.0, or dev-1.x branch)
Prerequisite
- [X] I have searched Issues and Discussions but cannot get the expected help.
- [X] I have read the documentation but cannot get the expected help.
- [X] The bug has not been fixed in the latest version.
Environment
Ubuntu 20.04, CUDA 12.2, cudnn 8.6.3。 2024年12月9日最新版本代码。
Describe the bug
一个简单的跳起、落下 动作，会被识别成4种动作分类：dancing gangnam style、jumpstyle dancing、tossing coin、high kick。
Reproduces the problem - code sample
1，执行命令：
python demo/webcam_demo.py configs/recognition/timesformer/timesformer_spaceOnly_8xb8-8x32x1-15e_kinetics400-rgb.py checkpoints/timesformer_spaceOnly_8xb8-8x32x1-15e_kinetics400-rgb_20220815-78f05367.pth tools/data/kinetics/label_map_k400.txt --average-size 5 --threshold 0.1
2，实际动作： 站在摄像头前，站立、曲膝、跳起、落下、站立，重复几次。 3，被识别成： dancing gangnam style: 15.33 jumpstyle dancing : 12.91 落地前被识别为 tossing coin: 16.63 有2次 有一只脚膝盖稍微外歪一点就被认为是 high kick : 11.86.
4，期待： 动作分类更规范，不与流行舞蹈挂钩，比如 跳起来 统一识别为 jump，而不是dance。
Reproduces the problem - command or script
python demo/webcam_demo.py configs/recognition/timesformer/timesformer_spaceOnly_8xb8-8x32x1-15e_kinetics400-rgb.py checkpoints/timesformer_spaceOnly_8xb8-8x32x1-15e_kinetics400-rgb_20220815-78f05367.pth tools/data/kinetics/label_map_k400.txt --average-size 5 --threshold 0.1
站在摄像头前，站立、曲膝、跳起、落下、站立，重复几次。
Reproduces the problem - error message
（没有错误信息，而是给的结果不正确）
Additional information
无。
我参考下面链接demo这里 来识别录像中原地跳的动作，发现一次都没成功。骨骼检测准确，但跳跃动作识别糟糕。看来依赖kinetics数据集来检测原地跳跃并不靠谱。 https://github.com/open-mmlab/mmaction2/blob/main/demo/README.md#video-structuralize-demo