Skip to content

detrex v0.4.0 Release
Compare
Choose a tag to compare
@rentainhe rentainhe released this 02 Jun 09:26
· 35 commits to main since this release
v0.4.0
85ceae2
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

Main updates

  • Support CO-MOT aims for End-to-End Multi-Object Tracking by Feng Yan.
  • Release DINO with optimized hyper-parameters which achieves 50.0 AP under 1x settings.
  • Release pretrained DINO based on InternImage, ConvNeXt-1K pretrained backbones.
  • Release Deformable-DETR-R50 pretrained weights.
  • Release DETA and better H-DETR pretrained weights: achieving 50.2 AP and 49.1 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
Assets 2
Loading