Wwupup
Wwupup
人脸关键点有助于loss更好的收敛，对人脸检测精度的提高很有帮助。就算不需要输出人脸关键点，也没必要在训练阶段进行删除。公布的精度就是直接由`yudacedet.yaml`配置文件得到的。另外更改测试时`score_thresh`和`nms_thresh`对结果会有影响
You can execute the following tool function： https://raw.githubusercontent.com/ShiqiYu/libfacedetection.train/master/tools/gt2coco.py to generate your own COCO-format annotation file.
Hello, I'm sorry to answer the question now. The `gt2coco.py` I mentioned earlier is a script that converts groundthruth to COCO format, where the segmentation tag is empty. When converting...
It seems that you want to get annotations about the facial key points of the WIDER-FACE dataset. That can be obtained by reading `trainset.json` directly.
In function https://github.com/ShiqiYu/libfacedetection.train/blob/a3bc97c7e85bb206c9feca97fbd541ce82cfa3a9/tools/gt2coco.py#L26 you can add the annotation include `bbox`, `keypoints `of your custom dataset to this data struct: ``` annos.append({ "segmentation": [], #your keypoints, for example: [[x1, y1, x2,...
Hello, this is the first time I submit code to MMDetection. I've checked the failure details but can't figure out what's wrong, please help me check if it's convenient.
> Hello @Wwupup. Thanks for your contributions. You can first follow the [CONTRIBUTING.md](https://github.com/open-mmlab/mmcv/blob/master/CONTRIBUTING.md) to install pre-commit and fix the lint error. Besides, it's better to also provide a config file...
你好，当前版本的人脸检测模型仅在WIDER-FACE数据集（比FDDB更难）上进行训练并测试，所以直接用这个权重测试FDDB数据集的话结果稍低是正常的。
> 为了兼容更多cpu 的指令集加速，acx2 一般是指2012 年后的cpu 了 ，在这之前的cpu 性能更差，感觉更需要这个，请问有什么办法能支持 SSE 指令集计算加速吗 目前没有针对其他指令集（除avx2/avx512/neon外）进行优化，你可以自行在facedetect_cnn.cpp中进行特定指令集的优化实现。
HelIo @YunYang1994 , have previously attempted to directly use focal loss as the classification loss, using the recommended hyperparameters from the paper, but the improvement seemed insignificant.