Accuracy is Not All You Need: Learning Consistent Video Segmentations
, Director of Research, Lightricks
When humans see segmentations — that is, when we get to see the segmentation mask, or some visual effect on top of it — we're actually much more sensitive to temporal consistency than to accuracy (beyond a certain threshold). I'll explain how we approach this problem — in data creation, in training our models, and in inference time — to create consistent masks of objects in videos.