Batching Helpers Demo

This demo shows how to use accvlab.batching_helpers to batch the matching and loss computations in StreamPETR.

See also

See the Evaluation for measured speedups, the experiment setup and commands to compare the baseline and optimized implementations.

The demo is structured as follows:

• Overview
• Batching Matching
  • Overview and Goals
  • Pre-processing
  • Cost Matrix Computation
  • Perform the Matching
  • Combine Results and Post-processing
  • Note on the 3D Assigner Implementation
• Batching Loss — Focal Head
  • Overview and Goals
  • Top-level Loss Computation Implementation
  • Detailed Loss Computation Implementation
• Batching Loss — StreamPETR Head
  • Overview and Goals
  • Top-level Loss Computation
  • Inside loss_batched()
  • Other Notable Implementations Changes
• Evaluation
  • Setup
  • Results
  • How to Run
• Summary