Yolox inference issue #452
Description
In the provided yolo-x.json file, it is mentioned that it should be used for YOLOX-based models, but there is no mention of preprocessing steps. In the official YOLOX repository, line no 242 https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/data/data_augment.py they apply letterboxing, with a padding value of 114 for all three color channels. However, the preprocessing details are not specified in your JSON. Additionally, the post-processing step, specifically the use of multi-class NMS, is also not mentioned. In the YOLOX repository, we can observe the use of multi-class NMS. My question regarding NMS is whether you are using class-agnostic or class-specific NMS in your pipeline.
We are currently using the following JSON (yolox_ours.json) for our inference. However, we are encountering false positives with very high confidence when comparing the results to the Python-based inference pipeline that uses OpenCV's VideoReader. Could you please help identify what might be causing this discrepancy?