Questions about different augmentors for sparse and dense gt.

[zhujiagang]

Thanks for sharing your excellent work! Why do you provide two different ways for augmenting data? The sparse one does not have yjitter，asymmetric color augmentation. Besides, why do you provide a new function resize_sparse_flow_map, instead of using cv2.resize(flow, interpolation=cv2.INTER_NEAREST)for resizing sparse gt?

[lahavlipson]

When altering sparse ground truth stereo, the mask of valid correspondences needs to be updated to reflect this change. This is what resize_sparse_flow_map does.

Other than that, the SparseFlowAugmentor and FlowAugmentor classes are pretty much the same, except the latter has a few extra features since the primary data we use for training (i.e. sceneflow) is dense.

[zhujiagang]

When altering sparse ground truth stereo, is there any difference if we use cv2.resize(flow, interpolation=cv2.INTER_NEAREST) andflow > 0 to set the valid flag, instead of resize_sparse_flow_map.

[lahavlipson]

Apologies, I didn't see your reply until now. When resizing the ground truth flow, you also need to scale the coordinates.

[zhujiagang]

Thanks a lot. Yes, we have resized the ground truth by cv2.resize(flow, interpolation=cv2.INTER_NEAREST) and scale.

[lahavlipson]

Closing this issue.