Video inference demo multi parallel inference optimization 视频推理demo多并行推理优化

[OpenMMLab-Assistant-004]

Video inference demo multi parallel inference optimization

视频推理demo多并行推理优化

In $MMPOSE/demo/topdowndemo_with_mmdet. py, when determining that the input format is video and the running device is GPU, data parallelism is used for inference. Specifically, this includes:

1. Treat each frame of the image as a single sample and preprocess the data corresponding to the target detection model in a multi-process manner
2. Package multiple frames of images into a batch and call the target detection model for inference
3. Treat each detected target box and corresponding image as a single sample, and preprocess the data corresponding to the pose estimation detection model using a multi-process approach
4. Package multiple targets into a batch and call the attitude estimation model for inference
5. Aggregate the pose estimation results of multiple targets on each frame of the image“

在 $MMPOSE/demo/topdown_demo_with_mmdet.py 中，判断输入格式为视频且运行设备为GPU时，采用数据并行的方式进行推理。具体包括：

1. 将每帧图片视为单个样本，用多进程的方式进行目标检测模型对应的数据预处理
2. 将多帧图片打包成 batch 并调用目标检测模型进行推理
3. 将检测到的每个目标框和对应图片视为单个样本，用多进程的方式进行姿态估计检测模型对应的数据预处理
4. 将多个目标打包成 batch 并调用姿态估计模型进行推理
5. 聚合每帧图片上多个目标的姿态估计结果

Task Information

Repository：MMPose

Difficulty：Medium

Points：30

Estimated Time Consumption：14 days

Task Status

Claimant's Github ID：None

Claim Time：None

PR Link：None

Points to Note

To claim the task, please reply Claim the mission in the comment section. After claiming the task, please add our assistant bot called MeowMewo on WeChat at: openmmlabwx.

After completing the task, please submit a PR with the title starting with [MMSIG]. After submission, please reply with your PR link in the comments section of this Discussion. We will review your work as soon as possible.

在评论区回复 Claim the mission 领取任务，领取任务后请及时添加小助手喵喵微信: openmmlabwx

完成任务后提交相关链接（如是PR，标题请以[MMSIG]开头）回复到本Disucssion评论区，我们会第一时间审核您的贡献。