Where is the dataset of YOLO algorithm?

[ai_curious]

@BryanEL here are some thoughts from my experience training YOLO v2 level code on a custom dataset that have been memorialized in other threads:

Lessons learned training YOLO from scratch on custom images

My YOLO v2 model operates, but it is not a very good learner. I spent time studying edge conditions in the data to see if I could figure out where noise might be creeping in. Here are a few things I discovered. Ground Truth for small objects causes issues Below is part of an image from the BDD training data I’ve mentioned in other posts. The first image shows ground truth in blue. The second image shows the same image with 32x32 pixel grid cells overlayed. Notice that the two grid cells …

Defining a custom class for YOLO loss

A fellow learner reached out and asked me about my YOLO implementation. Some of the code is a mess right now because I was trying to break it apart in order to be able to train the classification separately from localization (as suggested in one of the YOLO papers). The classification layers are a subset of the network architecture and produce a different output shape, then the localization is trained using transfer learning techniques but again, a different output shape. Pretty significant c…

Deriving YOLO anchor boxes

Deriving Anchor Boxes for YOLO One area of recurring questions about YOLO is anchor boxes, introduced with the second 2016 paper, YOLO9000 aka YOLO v2, as an optimization to help improve model stability during training. I’ll defer discussion of the role they play for now and just cover how they are created. Once you see what they are, I think it will be easier to understand how they are used. The Dataset I worked with the Berkeley DeepDrive 100K dataset (BDD). To keep the size manageable, I u…

Here’s a tldr

training YOLO from scratch is completely non-trivial. It takes a lot of data that you know well and a lot of computation. You need to do augmentation to ensure objects exist in all locations of the training set, otherwise you train predominantly on finding objects in the center of an image and most grid cell locations have no training at all.Grid cell and anchor box size needs to be tuned to the data, especially if you expect many objects small relative to image size. Speaking of anchor boxes, it is very important that you not use default anchor box count and shapes if those don’t match your dataset as it will lead to poor training outcomes. You will need to do the k-means analysis yourself to find the optimal number of and shapes of anchors for your data. Every object detection data set I looked at either used JSON or XML and every one had a complex record that included the bounding box and the class. If they are not paired in the training data, you can’t do detection (meaning localization + classification) until they are. The data set I worked with, mentioned in and linked from those threads, used JSON. Each record had lots of things I didn’t care about, so I had to write my own crawler to extract the bits I wanted. At least circa v2, which is what this course was based on, the YOLO creators broke the pipeline apart and had different architecture for location training and classification training because trying to learn all the parameters at once was too hard. My experience suggested I should do the same but I never completed the refactoring.

If you do go down this path, please post your experience.