Accuracy Ultralytics RT-DETR vs Pytorch models.

[rpatidar]

I was trying to do people detection and was experimenting on the models. attached in this repository and one with ultralytics-rtdetr

This are my observation

1. Ultralytics model perform much better in people detection vs the heaviest model(resent 101..). it misses multiple but ultralytics based model was able to detect majority of the people with high accuracy.
2. Problem with Ultralytics model i found was it was giving duplicate rectangle which I understand is a solved problem in RT-DETR.

-- any info on why this discrepancies are there ? have below questions :

1. are ultralytics one and one here are the same model or they are trained differently
2. is there any equivalent configuration for ultralytics model here that can be used for training ?
3. May be its a question belongs to ultralytics forum, why ultralytics baed model gives duplicated nested boxes which should not be happening with RT-DETR ?

[lyuwenyu]

ultralytics implements these two models https://github.com/lyuwenyu/RT-DETR/blob/main/rtdetr_paddle/configs/rtdetr/rtdetr_hgnetv2_l_6x_coco.yml https://github.com/lyuwenyu/RT-DETR/blob/main/rtdetr_paddle/configs/rtdetr/rtdetr_hgnetv2_x_6x_coco.yml