Issue with detecting high aspect ratio and small objects. about yolov6 HOT 6 OPEN

Hi @KaranBhuva22 , it sounds like an interesting project. As for the performance of trained model, I wonder if it is related to the data and label distribution of training set.

from yolov6.

Hi @Chilicyy,
I don't think it is related to data distribution issues because there was enough data for each class to fine-tune the model.

from yolov6.

Hi @KaranBhuva22 , did you do training with fuse_ab mode open？

from yolov6.

Yes @Chilicyy fuse_ab mode is open. I didn't modify it.

from yolov6.

@KaranBhuva22 Well, as for the problems you encountered, there's a few possible ways for improvements, I hope it helps.

1. With fuse_ab mode open, check anchors through AutoAnchor referring to the feature from yolov5, and then modify the initialization of anchors in config file.
2. Try to train from scratch instead of fine-tuning models, for the data distribution is rather different from COCO dataset.
3. Apply a larger scale of YOLOv6 model, such as yolov6n upgrade to yolov6s, to check if the model capacity is enough for your custom tasks.

from yolov6.

@KaranBhuva22 Well, as for the problems you encountered, there's a few possible ways for improvements, I hope it helps.

Hello @Chilicyy, Thanks for the quick response.

1. With fuse_ab mode open, check anchors through AutoAnchor referring to the feature from yolov5, and then modify the initialization of anchors in config file.

fuse_ab mode is not open since I am fine-tuning with yolov6l6_finetune.py, Anchor Aided Training Mode is currently unsupported on P6 models."

2. Try to train from scratch instead of fine-tuning models, for the data distribution is rather different from COCO dataset.

Which config file should I use for L6 model scratch training? Is there anything else I need to change except the configuration file for scratch training?

3. Apply a larger scale of YOLOv6 model, such as yolov6n upgrade to yolov6s, to check if the model capacity is enough for your custom tasks.

As I mentioned above, I am using the largest model, 'YOLOV6-L6', with 1280 resolution.

from yolov6.