Did you fine-tune a pre-trained YOLOv3 or train this from scratch?

[Karthik-Suresh93]

As you have only a few images here, I am curious if you trained on this dataset from scratch or fine-tuned a few layers in the darknet architecture using a pre-trained model? Could you please share your training strategy?

[jekhor]

I followed instructions at yolo author's site and use pretrained weights for convolutional layers as he described. чт, 19 апр. 2018 г., 10:12 Karthik-Suresh93 <[email protected]>:

…

As you have only a few images here, I am curious if you trained on this dataset from scratch or fine-tuned a few layers in the darknet architecture using a pre-trained model? Could you please share your training strategy? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAHK29QscrC0vDh9RvJuQ7bGujsmI0_Dks5tqDjegaJpZM4TbQuL> .

[Karthik-Suresh93]

But you have very few images to train on, it didn't produce any results for me. Thanks for your help anyway

[jekhor]

This dataset contains ~4000 labeled objects, this is not too small amount. But the net trained with this data doesn't show good results with photos from completely different series (taken with different device, weather conditions, zoom etc.) and you need additional samples in such case. I can recommend to take a look for openaerialmap.org for example (select filtering by high resolution).

[jekhor]

This is result from my current net applied to the San Francisco's photo from openaerialmap, for example.

[Karthik-Suresh93]

Could you provide a link to the full dataset, please? Your uploads contain about 130 images

[jekhor]

This is my full dataset. Every image contains number of cars (see .txt files) чт, 19 апр. 2018 г., 22:07 Karthik-Suresh93 <[email protected]>:

…

Could you provide a link to the full dataset, please? Your uploads contain about 130 images — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAHK2zdaW9QIDJiwHL8cqVoAVfFGw3JOks5tqOCEgaJpZM4TbQuL> .

[Karthik-Suresh93]

Did you freeze any weights or train the full model?

[jekhor]

Full train as described here: https://pjreddie.com/darknet/yolo/
This was my first attempt to get some results with YOLO, so I just repeated instructions written by its author.

[Karthik-Suresh93]

Thank you very much for your help!