Guidance on Labeling Obscured Objects for YOLOv5 Segmentation

[danielzhangau]

Hello YOLOv5 Maintainers,

I am seeking advice on the best practices for labeling partially obscured objects in instance segmentation using YOLOv5. I have encountered a scenario where an object is split into two visible parts due to occlusion by another object in the frame, and I am uncertain about the most effective way to label this for YOLOv5 segmentation.

Specific Scenario:

• An object (e.g., a piece of equipment) is partially hidden behind another, resulting in the object being visible in two separate parts.
• The challenge lies in how to label these parts for effective instance segmentation using YOLOv5.

Questions:

1. Labeling Approach: Is it more effective to label each visible part of the obscured object separately, or should we connect these parts with a thin line or minimal bridging annotation to indicate they are parts of the same object?
2. Impact on Model Performance: Which labeling approach would YOLOv5 handle more effectively for accurate segmentation? Are there specific considerations or adjustments to the training process recommended for such cases?

I would greatly appreciate any insights, recommendations, or references to documentation that might help in addressing this challenge. Your expert advice would be invaluable in ensuring the effectiveness of our segmentation model.

Thank you for your time and assistance.

Best regards,

[danielzhangau]

After I did some more research, I found this issue is related. But not tested yet. Does anyone know if this is working?