After attempting to train a custom model for object detection using two different video guides, I have decided to come to the forums, as I’m very stumped!
The goal that I am attempting to accomplish is to train a custom object detection model in order to detect when a goal has been scored in amateur football games. I also wish to detect individual players so that I can track their performance over the course of the video.
The first Youtube course video - Tensorflow Object Detection in 5 Hours with Python | Full Course with 3 Projects - YouTube - uses the Tensorflow 2 model zoo in order to train the model off a pre-trained model. When I get to the training section of the course, I execute the code in the command prompt, as directed by the video, however no progress happens beyond a certain point, and no errors flag up. This leaves me confused as to whether anything is happening in the background, but so far, after leaving it on my computer for a couple days there has still be no progress beyond this point.
Here is my code so far, using the [SSD MobileNet v2 320x320] as a preliminary test before I move on to a better detection model:
(Screenshot - 290012c8a4e13adff6a51e48917a7580 - Gyazo)
(Screenshot - ce4a5c12951f7adbb3687ee8773d24e7 - Gyazo)
(Screenshot - a3c332170d582fdfb3d05b67a2a0ee13 - Gyazo)
(Screenshot - 048c2a0a8cb8a52522e41821c8ec1aa5 - Gyazo)
When I input the command in the command prompt, I get this outputted:
(Screenshot - 2f95b900d924ce184a6a67e4b2c3126f - Gyazo)
(Screenshot - d2e4ecb7bd6a97a7303926eaab3ce8a0 - Gyazo)
I originally thought that this was just the beginning of the training and so left it to run, but after the few days of no progress, I closed it down. Upon reflection, I believe that my lack of progress was down to the difference in labelling my images. According to the tutorial, each image was to be hand-labelled, with one annotation file per image file, however, I used a tool called hasty.ai which implements auto-labelling, as my image bank was 200+ photos. Annotations were stored in two separate json files (one for training and one for validation), in COCO format, so I found a tutorial that trained a YoloV7 model on a custom COCO annotated images (therefore the same annotation format as my data).
Is my instinct correct here or could it be something else that I have overlooked? I am new to using Tensorflow and machine learning model training in general, so it could be the latter.
The second tutorial is here: YOLOv7 | Object Detection on Custom Dataset - YouTube
With my code and error here: (Screenshot - aff40a1a7de4d8c866c23a24616031e2 - Gyazo)
After extensive googling, I could not solve that ‘logger’ error, and so have come to you guys for (hopefully!) some help. I have tried re-installing alfred but to no luck, and have no clue where to go.
To make some progress with the project, I have implemented sound analysis on the videos, in order to at least detect when the crowd cheers, which gets pretty good detection rate on goals scored, however, this does still not let me track the players.
Another question of mine is it actually feasible to track individual players over a lengthy period of time, when they are moving across each other frequently? The sample videos aren’t in great quality and most of the time the players are super far away and very small on screen.
Any help or leads of where to get help would be greatly appreciated!
Thanks for taking the time to read this - I understand that this is a long post, but I think the detail is required.