Add Jersey Number Identification

[sahil-chaudhary]

The plan is to use jersey number to keep track of players and fuse it with processed data from DEEPSORT/BYTESORT.

The goal of the model is to track players by their jersey number. This would include three steps:

1. Jersey Number Identification (JN Identifier)
2. A good and light tracking algorithm to keep track of detected objects
3. Merging both JN identifier and Tracker

This should result into a video where the objects are labeled as {Team: {Team_ID}, JN: {int x}, type: $z \in {player,refere}$ }

[sahil-chaudhary]

Low accuracy due to excess neurons than in the output at the final layer. A generic framework has to be built rather than a simple resnet framework for training

[sahil-chaudhary]

The problem doesn't happen to be with the model but with the dataset. The dataset which claimed to have unique numbers from 1-99 in the jersey, did not satisfy the criteria and hence causing the problem. Now, it has to be corrected either by training the model to learn just these unique numbers or increase the dataset by introducing missed numbers in the dataset

[sahil-chaudhary]

The accuracy seems to have improved. The Integration was successful. But the problem on testing or any random football game remains. The solution is simple: We need larger dataset. Soccernet doesn't seem to be enough

[sahil-chaudhary]

https://github.com/user-attachments/assets/6f020ad7-df67-4806-aed8-dec7514e8ab6

The integration is flawless, but the results are disappointing. The next plan is to create more dataset to train it from scratch with manually annotated dataset.

[sahil-chaudhary]

Finally, the code works but another new problem pops up. Now, I finally understand why none of the youtubers has attempted this idea. The inference time is so high that it doesn't make any sense to integrate with the entire pipeline. The below shows the results for the model:

[sahil-chaudhary]

https://github.com/mkoshkina/jersey-number-pipeline/issues/4

I have attached the link to the issue raised in the original paper.

Few modifications and solutions are discussed in the discussions but there is only 2 days left for the project deadline, and it doesn't make any sense to put more time into this right now. Depending on the response, will continue the work done here.

But for now, I am leaving this issue open and will close accordingly.