How to obtain a good accuracy + how to get mAP?

[Robin2091]

Hi, I am training on a custom dataset with yolo tiny and 1 class. The model only provides an accuracy measure. Is there a way to obtain mAP? If not can anyone explain how I can create my own script to obtain mAP? I just noticed that mAP is the metric used in a lot of the deep learning object detection papers I have been reading so that is why I am interested in this.

Secondly, I trained on my custom dataset that has 3000 images. I used pre-trained weights and only got an accuracy of 30%. Is this normal? Also in the official darknet implementation, it said that loss should be less than 1. My validation loss was 12 and train loss was 5 after 30 epochs. However, my model pretty much plateaus after 30 epoch so I don't see much improvement. So is there something wrong with these numbers?

Also, I made some improvements with my dataset and got a 10 percent increase in accuracy when I tried it again(40% accuracy). So making improvements and removing bad data might help more but I'm not too sure. I need some confirmation related to whether the accuracy and loss values I got are normal for tiny.

• I loaded the pre-trained weights as described in issue #48 (it worked for me). I loaded the darknet weights and trained the top layers as described.

Thank you.

[zzh8829]

mAP calculation is pretty hard to implement, i have looked around on the internet hoping to find a clean algorithm i can borrow from. But unfortunately i haven't found a good implementation and also did not have time to write one from scratch. I definitely want to have mAP score myself too, but its just too hard to code.

[Robin2091]

Hi @zzh8829, I will try to implement a mAP score. I only have 1 class for training which might make it slightly easier. I am also thinking of implementing mAP post-training rather than during training.

I also found this repo, which might be helpful. https://github.com/Cartucho/mAP#explanation

Also, do you know how accurate the yolov3 tiny model can get? I have around 4000 images for training I only get around a 30 percent accuracy in tiny(around 50 percent in just v3). I think it's because I have bad data so I relabelled and got more images and I am going to train again. But I just want to know the general accuracy score for yolo tiny so I can compare.

Thank you, Robin

[zzh8829]

Please check out the new training tutorial for getting good accuracy

[bsarden]

It’s probably better to standardize on one implementation in order to avoid potential accuracy loss as a result of implementation details and to make the results more reproducible. We could probably just rely on the PyCOCO API to report the mAP as shown here:

https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb

This appears to be the same flow that Tensorflow/MLPerf relies on as well.

https://github.com/tensorflow/models/blob/master/research/object_detection/metrics/coco_tools.py

https://github.com/mlperf/inference/blob/master/v0.5/classification_and_detection/tools/accuracy-coco.py

[krxat]

@zzh8829 any update on this issue?? Thanks

[nelsoncardenas]

A good general option to measure the Mean Average Precision is using the development of João Cartucho (it uses txt files).

✔️ https://github.com/Cartucho/mAP

Also, the pip package from Simon Klimaschka based on Cartucho's work (it uses dictionaries maybe is useful for small datasets) ✔️ https://pypi.org/project/mapcalc/ ✔️ https://github.com/LeMuecke/mapcalc

I'll try this way to solve the problem in my project.

[aningineer]

For COCO dataset, perhaps you can use the pycocotools API: https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb