Hi,
I am a beginner.
i was working on a safety critical hand detection. i am using my own custom model(y11 instance seg). i was in a doubt whether i should add shadow part in a hand while annotation. in some images shadow of the machine is falling on the hand so i was confused whether i should add this image for training. Also i did a 100 epoch training but the mAP is coming around 73%. How can i improve my model accuracy? will this accuracy increase if i include shadow images for training? also what could be the best batch value and conf value for a safety critical object detection? please guide me on this.
Can you show examples of the annotated images with and without shadows you mean?
1 Like
Hello kiranraj, thanks for reaching out. For safety-critical applications like hand detection, achieving high robustness is crucial. Including training images with varied lighting conditions, including those where shadows fall on the hand, generally helps the model become more robust. When annotating, ensure you consistently and accurately segment the entire hand, even the parts covered by shadow.
To potentially improve your 73% mAP, consider augmenting your dataset with more diverse examples (like the shadow images), training for more epochs, or exploring hyperparameter tuning. You can find general advice in our Tips for Model Training guide.
The best batch size usually depends on your available GPU memory; experimenting is often necessary. For safety-critical systems, adjusting the confidence threshold (conf) is important. A higher threshold reduces false positives (incorrect detections) but might increase false negatives (missed hands). This trade-off between precision and recall needs careful evaluation based on your specific safety needs. Understanding metrics like precision and recall, discussed in our YOLO Performance Metrics guide, is key here.
1 Like
A very big Thank you for Burhan and pderrenger for the valuable guidance. it gave more insight to me. Thank you so much Team.
1 Like
You’re very welcome, kiranraj! We’re glad we could help. The Ultralytics team and the wider YOLO community are always happy to support users in their projects.
Good luck with your safety-critical hand detection application!
1 Like