incorrect detections for cars after fine-tuning yolov5l

[anisha150213]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

Hi,
I fine-tuned the yolov5l model with a custom dataset. Images are from paranoma video. I have two training classes: car and person and around 8k images, the number of instances of car is approximately 6k. The original image dimension is 1980x1080 and trained with 640x640.
I trained my model with 50, 100, 300, 400 and 500. With all the models, I faced the same issue. I used different iou-threshold in testing. But nothing worked that much. Can anyone please explain why fine-tuning is not providing better results?
Please take a look at the attachment for reference (screenshot).
reference_yolov5l.zip

Thanks

Additional

No response

[glenn-jocher]

@anisha150213 hi,

Thank you for your detailed report. Please ensure you are using the latest version of YOLOv5. If the issue persists, consider checking your dataset for class imbalance and ensuring your annotations are accurate. Additionally, experimenting with different augmentation techniques and hyperparameters might help improve your results.