search

enesozi / object-detection

Object detection on thermal images(FLIR dataset)
MIT License
90 stars 26 forks source link

Small, medium, large #5

Closed Pari-singh closed 4 years ago

Pari-singh commented 4 years ago

Hi @enesozi

You mentioned you sampled the large objects from the train set. Could you shed some light, how did you do that? I am training model on the same dataset using yolov3-spp and just training the final layers as I am short of resources to train from scratch. Getting not so good results.

enesozi commented 4 years ago

What do you mean by sampling the large objects? Oversampling the bicycle class? I didn't sample any objects based on their sizes.

Pari-singh commented 4 years ago

@enesozi , in your blog post, you mentioned you were getting high false positives since many of the small objects were not labeled in FLIR dataset and thus you removed those images. How did you remove them? Any routine to follow?

Thanks for your time!

enesozi commented 4 years ago

I didn't remove any images. I just didn't use small area' as a criterion in COCO evaluation. In other words, I wanted to see how better results I can get if I use only medium and large boxes from the ground truth annotations. If you would like to remove those images from your dataset, it'd require you to visualize ground truths on the images and take a look at them.

Pari-singh commented 4 years ago

Oh I see, so you did the training as usual and then only while evaluating your model, you gave that criteria? Sorry, i wasn't aware of such options in COCO evaluation!

Thanks a lot!

enesozi commented 4 years ago

Yes, exactly. You may refer to the official ipynb to see how it works.

Below code snippet is an example for excluding small area from the evaluation. cocoEval.params.areaRng = [[32 ** 2, 1e5 ** 2],[32 ** 2, 96 ** 2], [96 ** 2, 1e5 ** 2]] cocoEval.params.areaRngLbl = ['all', 'medium', 'large']