I was excited about the D-FINE model, but I have got ABYSMAL Results

🤗Transformers
1

Hi all,
As the title says, I have been super excited about the D-FINE Object detection model that HF released with the hope to replace YOLO & Detectron2 that I have been using.

I have followed this example notebook by @qubvel-hf and I have customized it for a zindi challenge I’m participating in.

But I have gotten abysmal results with it with very bad mAP results.
The only DFINE model I was able to run in Colab/kaggle was the “ustc-community/dfine-nano-coco”. Whenever I tried running the dfine small/large/x-large models, I would get CUDA OutofMemory within the first epoch.
For me, DFINE performed way worse than YOLO or Detectron2 and model training took 3+ hours when training for only 15 epochs.

Here’s my notebook

Here a few screenshots showing the poor results I got.

image
image690Ă—488 24 KB

mAP 75 is as low as 0.029600 after 15 epochs. This is very low, I have never seen that. YOLO or Detectron2 easily give me 0.5 or 0.6 with the same 15 epochs.

My submission to the Zindi challenge is appalling low when using D-FINE. See here

image
image1300Ă—626 50.2 KB

What am I doing wrong? Have you guys seen similar results with D-FINE?

Kindly help. @qubvel-hf

2

I’ve tried to finetune several dfines as well, all of them have terrible performances

3

Same here, so weird

4

Hmm… How about with official implementation?

5

did not try fine tunning DFINE yet, howver no issue running their pre-trained models with tensorrt or just with pytorch.

Accuracy wise, it performs ok for clean images, not so great on video

(blurry/compressed) frames but it does okish. A2.0 Licence are a great pros for it though, while Yolo etc licences are less flexible I think. just my 2 cents here.