Why this fork gives extremely low mAP and recall results

[novate]

Recently I trained a dataset of ~8000 images for ~50000 iterations on darknet yolov3(pjreddie's repo). Then I use detector recall(set the threshold at 0.5 in detector.c) and get a result of around 50%. Then I want to calculate mAP and checked this repo, using detector map and I get extremely low mAP of less than 1% with recall lesser. Then I tried detector recall on this repo and found that only 2 objects were detected all the round so the recall became less and less. Could anyone help me out?

P.S. My labels are directly yolo format, because I labelled them on LabelImg with yolo format. So that there is neither PASCAL format labels, nor label conversion is used.

If something more should be stated, please ask me to do so.

P.S.2 My dataset is with only 1 class: 'person'.

This is the result from pjreddie's repo. This is the result from AlexeyAB's repo.

[AlexeyAB]

I use detector recall

• You got IoU = -nan in the pjreddie's repo so it looks like very bad.

• Can you detect anything by using your cfg/weights file on the pjreddie's repo? ./darknet detector test...

• Check that you use the correct cfg and weights files.

• What params do you use in the Makefile?

[novate]

Thank you for your reply.

My conditions:

• IoU in log file is not nan, IoU in ./darknet detector recall(pjreddie's repo, yolov3) is nan, that may be because that there is bug in pjreddie's repo. https://github.com/pjreddie/darknet/pull/952

• Yes, I can detect something by using my cfg/weights file on the pjreddie's repo ./darknet detector test.

• Yes, I used the correct cfg/weights file, and I used darknet53.conv.74 for training.

• I set GPU=1, OPENCV=1 and else not edited.

P.S. I used AlexeyAB's repo to train with the same cfg/weights and darknet53.conv.74 and the same pics, same labels(everything is the same), but I cannot train properly as with pjreddie's repo(yolov3) did, everything goes to nan in the log file when training with AlexeyAB's repo.

This is some of the log file for training with pjreddie's repo.

This is some of the log file for training with AlexeyAB's repo.

[AlexeyAB]

• Can you share your cfg-file, and obj.data file?

• What command do you use for training?

• Can you show your files bad.list and bad_label.list?

• Do you use the latest version of this repository? https://github.com/AlexeyAB/darknet

• What CUDA, cuDNN and GPU do you use?

[novate]

• I don't have bad.list.

• Here is my command: ./darknet detector train EP.data yolov3-EP.cfg darknet53.conv.74 -gpus 1,2,3 > log_data.log

• I used the latest version of this repo.

• I used CUDA 9.0, my cudnn is broken on my server and I'm asking the admin to fix it, so I set CUDNN=0. I used 4 nvidia 1080Ti GPUs

• Here is a sample of my bad_label.list.

• My cfg-file and obj.data is in this .zip file. cfg_data.zip

[AlexeyAB]

• Here is a sample of my bad_label.list.

• Show several txt-files with your labels.

• In your labels (txt-files) should be only class_id=0: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

You should label each object on images from your dataset. Use this visual GUI-software for marking bounded boxes of objects and generating annotation files for Yolo v2 & v3: https://github.com/AlexeyAB/Yolo_mark

It will create .txt-file for each .jpg-image-file - in the same directory and with the same name, but with .txt-extension, and put to file: object number and object coordinates on this image, for each object in new line:

Where:

- integer object number from 0 to (classes-1)