Adaptive Spatial Anchor Management for Non-Cyclic Trajectories
Implement a dynamic spatial keyframe policy that replaces static anchors when view overlap drops, maintaining global consistency in long trajectories.
Source paper
Shuai Yuan, Yantai Yang, Xiaotian Yang, Xupeng Zhang, Zhonghao Zhao, Lingming Zhang, Zhipeng Zhang
Generated by AI. The paper’s authors did not write these.
Implement a dynamic spatial keyframe policy that replaces static anchors when view overlap drops, maintaining global consistency in long trajectories.
Augment token pruning with a spatial density score to ensure rare 3D regions retain enough points for complete mesh reconstruction.
Train a lightweight Multi-Layer Perceptron to predict token importance scores from keys and query statistics, avoiding expensive full attention matrices.
This paper introduces InfiniteVGGT, a transformer-based model for continuous 3D geometry reconstruction from infinite video streams without memory overflow. The authors propose a rolling memory mechanism that dynamically prunes the KV cache based on key-diversity (cosine similarity) rather than attention weights, enabling compatibility with FlashAttention and long-term stability.
The model utilizes a causal video transformer with a 'rolling memory' KV cache. To prevent memory growth, it employs a training-free pruning strategy that retains a set of 'immutable anchor tokens' (from the first frame) and selects the most diverse recent tokens based on a layer-wise adaptive budget determined by the cosine similarity of keys.