search

yoxu515 / aot-benchmark

An efficient modular implementation of Associating Objects with Transformers for Video Object Segmentation in PyTorch
BSD 3-Clause "New" or "Revised" License
583 stars 106 forks source link

Object merging in tracking #66

Open leonid-pishchulin opened 9 months ago

leonid-pishchulin commented 9 months ago

Hi, do you have any suggestions how to avoid grouping multiple spatially-separated objects into one during tracking?

leonid-pishchulin commented 9 months ago

example:

  1. final output from frame t-1 contains one object labels o1
  2. run tracking for frame t. Tracking outputs 2 disjoint objects with with the same label o1 How to update the code to make sure tracking doesn't output multiple objects under the same label?
z-x-yang commented 9 months ago

Commonly, it's reasonable to have multiple spatially separated objects with the same ID when an object is partially occluded and separated.

z-x-yang commented 9 months ago

If you want to keep only one object from multiple multiple spatially separated objects, a possible solution is to remove the object regions with lower confidence, which can be calculated by the confidence sum of all the region pixels.