RT-DETR in TAO: Extremely Low KITTI Scores (0.000–0.005) Despite Good Detections

AI & Data Science Computer Vision & Image Processing
1

I have trained a number of detection models using TAO Toolkit with (pretrained) RT-DETR.

NVIDIA Release 6.0.0-PyTorch (build 134983853)
TAO Toolkit Version 6.0.0

I only have one target class (plus the background). My targets are small relative to the image frame.

For training I had set num_queries to 300, and num_select to 10. (I ran a number of experiments changing these numbers; I need num_queries to remain high.)

When I run inference on the trained models- the output is in KITTI format. My KITTI scores are consistently extremely low- ranging from 0.0 to typically only 0.005 or thereabouts. However, I am getting good detections as evidenced by well positioned bounding boxes. I get several bounding boxes on the target with very low scores.

So- I don’t understand why my KITTI prediction scores are so excessively low…

2

Continuing the discussion from RT-DETR in TAO: Extremely Low KITTI Scores (0.000–0.005) Despite Good Detections:

I’ll add on- I also changed num_select back to 300- that helps the number of detections. However, the problem still remains- extremely low prediction values on inference; generally on the order of a maximum of 0.01. Often less.