MASt3R-SLAM Loop Closure – L1 vs L2 Similarity

I implemented and benchmarked an alternative L1 (Manhattan) similarity measure within the ASMK retrieval kernel. This involved:

• Kernel Modification: Swapping the dot-product calculation for a sum-of-absolute-differences operation, while preserving descriptor normalization and aggregation logic.
  
  

• Automation Pipeline: Extending the benchmarking scripts to run both L1 and L2 variants seamlessly across nine TUM RGB-D sequences, capturing Absolute Trajectory Error (ATE) and frame-rate metrics.
  
  

• Quantitative Analysis: Developing post-processing scripts to align estimated trajectories (via Horn’s method) and compute per-sequence ATE, then plotting paired accuracy vs. throughput comparisons.

• Accuracy Parity: Across all tested sequences, the L1-based loop-closure matched L2’s ATE within millimeter-scale differences, confirming negligible impact on localization quality.
  
  

• Performance Trade-off: Despite its algorithmic simplicity, the L1 variant ran 25–30% slower than the hardware-optimized L2 implementation—highlighting the benefits of fused-multiply-add acceleration for Euclidean computations on modern CPUs.
  
  

• Practical Takeaway: While Manhattan distance can serve as a drop-in replacement when hardware constraints limit vector-dot products, Euclidean metrics remain preferable for real-time loop-closure in high-throughput robotic applications.