AIBRIEF
F1 is a 0 to 1 score averaged across five object categories, and the same trained neural network on the public RadarScenes dataset of radar returns from real world driving scenes scored 0.38 at one radar point per tracked object and 0.76 at five.
A radar perception engineer tested whether the model or the data path set the ceiling on radar-only object recognition. The answer, on a public radar dataset, was the data. At low point density a stationary two-wheeler was misread as a pedestrian; at higher point density the error stopped.
Bruno Pinto trained a small 3-layer neural network (MLP) to recognize five categories (car, large vehicle, two-wheeler, pedestrian, pedestrian group) using only radar returns from RadarScenes, as posted on r/MachineLearning. The encoding followed the histogram recipe of Tatarchenko and Rambach: 16 bins each over radar cross-section, ego-motion-compensated velocity, and x/y position, plus an unbinned Doppler-spread feature.
With the same trained model and no other change, macro F1 climbed from 0.381 at one radar point per object to 0.764 at five. Large-vehicle F1 went from 0.037 at one point to 0.995 above eleven. Wider and deeper networks, six alternative feature encodings, and different bin edges all moved the score within a six-fold cross-validation noise floor.
Count returns per tracked object in your own data path before spending on bigger architectures. The honest limit is one engineer on one naturalistically imbalanced dataset with merged rare classes. Reproduce the check before declaring architecture tuning dead.