search

pjreddie / darknet

Convolutional Neural Networks
http://pjreddie.com/darknet/
Other
25.86k stars 21.33k forks source link

Anybody can reproduce the mAP reported in the paper on COCO dataset? #1362

Open muye5 opened 5 years ago

muye5 commented 5 years ago

I tried, but got a lower mAP_50，so anybody can share some results or messages.

duohappy commented 5 years ago

I just get about 45%, and I dont know how to improve it.

ghost commented 5 years ago

i have never seen the results of official weights when retraining. Even if i retrain with less categories

wonchulSon commented 5 years ago

How did you get the mAP? Plz!!! let me know TY

AlexeyAB commented 5 years ago

@wonchulSon

How to get mAP on MS COCO dataset Test or Validation: https://github.com/AlexeyAB/darknet/issues/2145

The simplest way is to use this repository https://github.com/AlexeyAB/darknet with this command ./darknet detector map cfg/coco.data cfg/yolov3.cfg yolov3.weights -iou_thresh 0.50

duohappy commented 5 years ago

@MartyEz what is your configure? I try to use 4 GPU to train the model, I finally can not get desired result

duohappy commented 5 years ago

@wonchulSon I use the repo, which AlexeyAB fork

duohappy commented 5 years ago

@AlexeyAB hi, Can you acquire the official result? I try many times, but I can not get the result

Aminkhormali commented 5 years ago

Dear @AlexeyAB, @duohappy , can you please let me know if you were able to achieve the same results as official paper through retraining YOLO V3 on coco dataset?

duohappy commented 5 years ago

hi @Aminkhormali ,I can not get the official result yet.

eric-tc commented 4 years ago

Hi @duohappy ,

any news about YOLO v3. I get the same as you 45%. Did you find a way to improve the result?

aningineer commented 3 years ago

Has anyone reproduced the original Yolov3 results using a minimal pytorch / tensorflow repo? Could you please point me to it?

I managed to get decent results using the Ultralytics repo (https://github.com/ultralytics/yolov3) but it's quite heavy to develop on.