Baseline
Removing the dropout from slot attention head improves performance
Baseline fails at identifying smaller objects and have blur boundaries.
Remove Dropout
Removing the dropout from slot attention head improves performance
Background/Foreground Aware Strategy
Identifying the background first helps to separate objects from the background.
Geometric Consistency Loss
Encourages slots to align with consistent 3D structure across frames
Leads to spatially coherent activations, but introduces training instability.
Deeper Slot Head
- Performance degrades as the number of objects increases, but adopting a deeper head improves robustness in multi-object scenarios.
- The decoder with a deeper head learns meaningful semantic cues, as evidenced by improved performance on the KITTI.