Hi, we are using the released subject-fidelity classifiers (pedestrian_final.pth, vehicle_final.pth, and we've also tried pedestrian_classification.pth / pedestrian_classificationv2.pth) for evaluation.
The issue we are encountering is that the pedestrian checkpoints seem to assign very high confidence scores to almost everything. For example, even crops containing cars or trucks can receive confidence scores above 90%. This behavior is very consistent across all three pedestrian checkpoints we tested.
We followed the evaluation pipeline in the official code. Specifically, the crops are preprocessed as:
PIL RGB → Resize([256, 128]) → Normalize(0.5, 0.5, 0.5)
and the confidence score is computed directly from the raw sigmoid output.
We also observe the same issue with the vehicle checkpoint.
Could you please let us know whether we might be missing any preprocessing, post-processing, thresholding, or other evaluation details when using the pedestrian classifiers?
Hi, we are using the released subject-fidelity classifiers (
pedestrian_final.pth,vehicle_final.pth, and we've also triedpedestrian_classification.pth/pedestrian_classificationv2.pth) for evaluation.The issue we are encountering is that the pedestrian checkpoints seem to assign very high confidence scores to almost everything. For example, even crops containing cars or trucks can receive confidence scores above 90%. This behavior is very consistent across all three pedestrian checkpoints we tested.
We followed the evaluation pipeline in the official code. Specifically, the crops are preprocessed as:
PIL RGB → Resize([256, 128]) → Normalize(0.5, 0.5, 0.5)
and the confidence score is computed directly from the raw sigmoid output.
We also observe the same issue with the vehicle checkpoint.
Could you please let us know whether we might be missing any preprocessing, post-processing, thresholding, or other evaluation details when using the pedestrian classifiers?