- Expand Metrics: Calculate additional quantitative performance benchmarks.
- Optimize Streaming: Implement memory caching to enhance throughput.
- Broaden Data: Increase the variety and volume of evaluation datasets.
- Refine Slot-Dropping: Experiment with alternative slot-drop strategies.
- Deeper Slot-Head: Explore deeper and more expressive slot head architectures to improve representation capacity.
- Mask-Guided Tracking: Leverage the learned segmentation mask to guide and refine the target tracking process.
- Object level supervision: Incorporate weak or pseudo supervision (e.g., masks from foundation models) to guide slots toward semantic object decomposition.
Reference
- Wang et al., “VGGT: View-Graph Generalizable Training,” CVPR 2024.
- D. Zhuo et al., “Streaming 4D Visual Geometry Transformer,” ICLR 2026.
- Xiao et al., “SpatialTracker: Tracking Any 2D Pixels in 3D Space,” CVPR 2024.