Questions on Football Players Tracking

[lpc-eol]

Hi there! Could you share some hints on the training settings for yolov5x 's best.pth? Like how did you modify the model structure of yolov5? And also, what dataset you used to train? Only the 600 images? The video is awesome and motivates me to do better!

In fact, I did apply yolov7 and the Bytetrack to this task. To achieve better detection of football, I try to train the model of yolov7e6e on an open dataset (Soccernet Tracking Dataset) and load the pre-train weight of yolov7e6e provided by the author with an initial training setting in 100 epochs. The input resolution is 1280. I found that the mAP@.5:.95 is only 0.42, which is very low—especially the detection accuracy of football. The training progress is shown in the following image.

Thank you for your assistance.

[SkalskiP]

Hi, @lpc-eol 👋🏻! I just noticed that you tried to reach out to me over email. GitHub issues or discussions are a better choice. I tend to be much more responsive here. I basically live on GH.

I'm not sure if you've had a chance to read my blog post, but I described in a little more depth the things I only mentioned in the video.

Like how did you modify the model structure of yolov5?

The architecture of the model is unchanged. I just trained my own model.

And also, what dataset you used to train? Only the 600 images?

Yes! And I tried to minimize the amount of manual work. So I pre-annotated images using the YOLOv5 COCO model. And I augmented annotated images. So, I only do annotation refinement, not full annotation.

The video is awesome and motivates me to do better!

You made my day with that comment!

Also, as a general comment, I found input resolution to be a key model hyper parameter when detecting small objects like football.

[SkalskiP]

You have done an excellent job so far! I like it a lot! I assume you used my dataset. Did you use any augmentations? What is the input resolution? What is the size of the model?

[angineerdev]

You have done an excellent job so far! I like it a lot!

@SkalskiP It is nice to hear that, thanks 😄

I assume you used my dataset. Did you use any augmentations?

I used version 2 of the dataset you provided at the following link: https://universe.roboflow.com/roboflow-jvuqo/football-players-detection-3zvbc/dataset/2. I didn't apply any additional augmentations apart from the ones already present in version 2.

What is the input resolution? What is the size of the model?

Both training and inference were done at 1280p. Used yolov5x6 as the base model.

To be clear, the training command I used, as mentioned in my previous comments, is as follows:
python train.py --imgsz 1280 --batch 12 --epochs 100 --data datasets/football-players-detection-2/data.yaml --weights yolov5x6.pt

[SkalskiP]

I'm pretty sure I used a higher resolution than 1280p.

[angineerdev]

Hah, no worries! I trained for 100 epochs and used 1280 as image resolution. I used the default learning rate. Let me know how it went and how can I help :)

Ah, based on this comment, I didn't attempt to go beyond 1280. I guess you were referring to a different training; my apologies! 😅
Then, let me try higher resolutions. Thank you so much for your time and help.

[SkalskiP]

I used 1280 for training, but I think I used more than that during inference. 🤔 I'm not super sure... Eh. That is why I need to start using W&B for my experiments.