- Offline data processing
- write an independent function, which inputs 2D keypoint, heatmap size and generates flux map and distance map.
- save the heatmap with correspond id so that we can load it during the init of dataloader
- Validation indicators
- we may use 2D distant (EPE) in the original images. In this case, we need a function to project our prediction back to the original image coordinate. I attached my function in the folder named for inference.
-
Store the model regularly during training
-
During the training of the network, try to explore the related papers. We can discuss the papers during the meeting.
- One we may expore is to build the connection between mask and flux.
- Once we have some initial results, we can further refine the pipeline/output maps or add more losses based on the analysis of the resutls.