Kudos on 6.3

[knoppmyth]

Before Asking

• [X] I have read the README carefully. 我已经仔细阅读了README上的操作指引。

• [X] I want to train my custom dataset, and I have read the tutorials for training your custom data carefully and organize my dataset correctly; (FYI: We recommand you to apply the config files of xx_finetune.py.) 我想训练自定义数据集，我已经仔细阅读了训练自定义数据的教程，以及按照正确的目录结构存放数据集。（FYI: 我们推荐使用xx_finetune.py等配置文件训练自定义数据集。）

• [X] I have pulled the latest code of main branch to run again and the problem still existed. 我已经拉取了主分支上最新的代码，重新运行之后，问题仍不能解决。

Search before asking

• [X] I have searched the YOLOv6 issues and found no similar questions.

Question

Kudos on 6.3

Additional

Kudos for the work on 6.3. I first built a custom object detector with 6.0 and it was state of the art (compared against Faster RCNN, YOLOv3 - 7)! It was even better than YOLOv8 for my dataset. I tried 6.2 and it wasn't as good as 6.0. In fact, for my dataset, it was worse. I'm however glad to say, 6.3 improved upon what I saw with 6.0. The only problem I see, is occasional false positives that I don't see with 6.0.

Thanks again for the great work!

[Chilicyy]

@knoppmyth Thanks for your attention, and we appreciate you saying that. As you mentioned above, the problem of v3.0 in your cases is occasional false positives. Could you show us more details?

[knoppmyth]

@Chilicyy You're welcome. The custom object I'm detecting, is in a product I'm hoping to commercialize so, I'm keeping it under wraps at the moment. I'll attempt to duplicate the issue with another dataset. If I can do that, I'll post about the details. This was more of kudos than a complaint. :-)

[knoppmyth]

@Chilicyy Where is an example of what I'm seeing: I use {(round(np.amax(list_of_conf) * 100) -1)} to give a hint of that the confidence should be set as if nothing is detected. I'll have to take a closer looking to see why it is saying "99" when my confidence is set lower.

Both models were built with the same dataset and trained for 50 epoches each. I export to ONNX for use with OpenCV. I've uploaded the dataset, args.yaml, etc., Google Drive. Please let me know when you've retrieved so I can delete the tarball.

[Chilicyy]

Hi, @knoppmyth It seems that you are comparing the performance of v3.0 model with v1.0 model. You could enlarge the confidence threshold for v3.0 model, because IOU branch was abandoned since v2.0, and the predict scores were higher than v1.0 models.

[knoppmyth]

Hi @Chilicyy , Yes, I'm comparing the two. I wanted to see how different models compared against once another. I did notice that v3.0 gives a higher predicted score! This is awesome but from what I've seen in v3.0, false negatives occur when they weren't present in v1.0.