modification of Yolo v2

[MyVanitar]

Hello Felix,

in the article of "Detect and Classify Species of Fish from Fishing Vessels ..." you mentioned:

I used an improved version of YOLO called YOLO v2. It made several small but important changes inspired by Faster R-CNN, such as assigning bounding box coordinate “priors” to each partitioned region and replacing the fully connected layers with convolutional layers, hence making the network fully convolutional. Essentially, YOLO v2 works very similarly to RPN.

May I ask you where you have made this modifications? inside the yolo-voc.cfg file isn't it?

I want to follow you and make these modifications. Would you please help me?

[flyyufelix]

Hello, I used the original implementation in Darknet (https://pjreddie.com/darknet/yolo/). You can follow this guide: http://guanghan.info/blog/en/my-works/train-yolo/ to train YOLO with your own dataset. I also hack into the C code to get the program output predicted bounding boxes.

[MyVanitar]

Thank you.

The article dated at 15/12/22. seems to be for YOLO V1

[MyVanitar]

Besides, Do you think YOLO v2 or SSD-300 (7++12+COCO) will have less detection errors? in Fine-Tuning.

Although I think we can not Fine-Tune the YOLO. The provided weights is as an initial classification weights and we train from scratch actually.

[flyyufelix]

Yes, the article is slightly outdated. It is for YOLO v1. However, the config files of YOLO v2 and v1 are very similar so long as I recall, so if you can follow that of v1 I am sure you can do the same for v2.

YOLO v2 should outperform SSD-300. Resolution plays an important part in detector performance (can refer to this paper: https://arxiv.org/pdf/1611.10012.pdf). YOLO v2 takes 448x448 inputs while SSD takes 300x300. Speaking from my personal experience, YOLO v2 outperformed SSD-300 in the kaggle competition. In this regard, YOLO v2 vs SSD-512 should be a more fair comparison.

Yes, the provided weights for YOLO v2 (DarkNet19 with 448x448 input resolution pre-trained on ImageNet) is for classification. But it's still very useful as good feature extractor for the detection layers. I used that for my kaggle dataset and it worked quite well.

[MyVanitar]

one missing feature in Yolo v2 is that they have not implemented the angle in training, it is a useful feature when the target objects might rotate in practice, which of course does not apply for the VOC dataset, and that's the reason why they have not implemented it completely.