2.99
Machine perception systems
Comma Video Compression Challenge
Metric-aware compression experiment
Developed and evaluated metric-aware video-compression approaches for machine-perception workloads, reducing the official evaluation score from 2.99 to 1.3546 while keeping the final archive below 1 MB.
Role and scope
Developed baselines, evaluated variants, validated CPU evaluation behavior, tracked side-channel size, and documented metric tradeoffs.
PythonVideo CompressionMetric EvaluationPoseNetSegNet
1.3546
Final score
994,341 bytes
Archive size
4,808 bytes
Side channel
Case study
What this shows
This project explored video compression through the lens of downstream machine-perception metrics and strict payload constraints.
Problem
The challenge rewards perception-aware compression rather than visual fidelity alone, so the work focused on model metrics, reproducible validation, and strict archive-size constraints.
Baseline
The baseline official evaluation score was 2.99 before compression and side-channel iterations.
Metric-aware approach
The work focused on reducing PoseNet and SegNet metric error while keeping the final archive under 1 MB.
Final validation
The CPU-validated final score was 1.3545992946, with a 994,341-byte archive, 0.00286877 PoseNet metric, and 0.00523134 SegNet metric.
Documentation discipline
The MPS sharp variant was documented separately because its CPU evaluation crashed, so it was not overstated as the final validated result.
Interested in the implementation?