pose proposal / ppn-resnet50-V2-HW results
korejan opened this issue · comments
I just briefly tried the pose proposal implementation, post-processing is significantly faster for sure, in my brief tests with a single web camera and batch size of 1 it took like ~0.065ms per frame. The model ppn-resnet50-V2-HW also uses less GPU and inference took ~3.5ms per frame (where as TinyVGG-V1-HW takes ~6.3 ms).
The only problem is that the accuracy/results using ppn-resnet50-V2-HW seems pretty bad compared to using TinyVGG-V1-HW/paf, I'm assuming this is to do with the model rather than any potential issues with the current PP impl? is there another PP model with higher accuracy? or one planned to be made? or a method of converting the other models?
This implementation may work but without a promise of accuracy. We are actively trying to improve the models' accuracy, and we will update models with higher mAP as soon as possible.
As far as I'm concerned, this method is suitable for scenarios with fewer people. Also, it is reported to be of bad accuracy on COCO dataset...
