Point density, not architecture, doubled radar classifier F1 from 0.381 to 0.764
bruno_pinto90 · reddit · 2026-09-07
A perception/radar engineer trained a radar-only 5-class classifier (car, largevehicle, twowheeler, pedestrian, pedestriangroup) on RadarScenes and found a counterintuitive result: point density, not model architecture, was the real bottleneck.
- Going from 1 to 5 points per instance roughly doubled macro F1 (0.381 → 0.764) with no model changes
- A single point can't carry size/velocity-spread signatures: largevehicle F1 is 0.037 at n=1 vs 0.995 at n=11+; twowheeler and pedestrian collapse to the same near-zero-velocity signature when sparse
- Ablations — wider/deeper nets, six feature encodings, different bin edges — all landed inside a 6-fold split sensitivity noise floor; explicit per-instance statistics actually scored slightly worse (0.658 vs 0.686)
- Data caveats: class imbalance, twowheeler merging two speed regimes, and a single stationary two-wheeler track dominating one sequence's distribution, explaining cross-fold variance
- Real failure case: a nearly stationary two-wheeler with one radar point predicted as pedestrian; a car in the same scene classifies fine on RCS/Doppler alone
Full writeup is open-sourced on GitHub; the author asks for sparse-scene representations and better ways to confirm real improvements.
More from Research
- AI-generated 20 kN methalox engine goes from spec to hot-fire in about three weeks — creatoroff · 2026-09-07
- RandKV ships as pip-installable random KV-cache eviction for Transformers, reports honest negative perf results — atease01 · 2026-09-07
- New essay explores the strange and marvelous challenges facing biological AI — anshulkundaje · 2026-09-07
- MIT and EPFL build a sub-300g robot that swims underwater then flies out — lukas_m_ziegler · 2026-09-07
- Interactive Speculative Decoding Tutorial for NeurIPS Explains When It Stays Lossless — Madisonkanna · 2026-09-07
- LLMs as a Cognitive Virus: New Paper Models AI Dependence Tipping Points — serrjoa · 2026-09-07