detrex v0.4.0 Release
Main updates
- Support CO-MOT aims for End-to-End Multi-Object Tracking by Feng Yan.
- Release
DINO
with optimized hyper-parameters which achieves50.0 AP
under 1x settings. - Release pretrained DINO based on
InternImage
,ConvNeXt-1K pretrained
backbones. - Release
Deformable-DETR-R50
pretrained weights. - Release
DETA
and betterH-DETR
pretrained weights: achieving50.2 AP
and49.1 AP
respectively.
Pretrained Model
DETA-R50-5scale-12ep
bs=8:50.0AP
DETA-R50-5scale-12ep
aligned hyper-param:49.9AP
DETA-R50-5scale-12ep
with only freeze the stem of backbone:50.2AP
H-Deformable-DETR-two-stage-R50-12ep
aligned optimizer hyper-params:49.1AP
DINO-R50-4scale-12ep
aligned optimizer hyper-params:49.4AP
DINO-Focal-3level-4scale-36ep
:58.3AP
Benchmark ConvNeXt on DINO
- convnext-tiny-384:
52.4AP
- convnext-small-384:
54.2AP
- convnext-base-384:
55.1AP
- convnext-large-384:
55.5AP
Benchmark InternImage on DINO
- internimage-tiny:
52.3AP
- internimage-small:
53.5AP
- internimage-base:
54.7AP
- internimage-large:
57.0AP
Benchmark FocalNet on DINO
- focalnet-tiny
- focalnet-small
- focalnet-base
Other pre-trained weights
- Deformable-DETR R50:
44.9 AP
(better than 44.5AP from original repo) Group-DETR-R50-12ep
:37.8AP