Mask_rcnn: The prediction class is always 1

[wadmes]

I used this work (just a mock of coco.py)and fed my own medical image data into the model,
the train and validation process looks good with a valid loss,
but the test result is terrible, and the class is always 1 with a high confidence.

Anyone has ideas to how to debug? I don't know how to see the result when training, which only indicates the losses.

[5730279821-TA]

1. Are you sure your model isn't overfit ? (check your mrcnn_class_loss and val_mrcnn_class_loss)
2. try to check your data whether each class data is balance.
3. try to edit DETECTION_MAX_INSTANCES and DETECTION_MIN_CONFIDENCE

[wadmes]

Thanks for your reply.

1. No actually. And both losses are around 0.4.
2. No, it is around 1:1.
3. I haven't tried it and I will have a try.

[waleedka]

Start by verifying that your Dataset class is generating the data correctly. Use the inspect_data notebook to visualize the images and verify that the code is generating the correct training targets.
If that's all okay, then use the inspect_model notebook to visualize the prediction process step by step and that should help you locate where the error is happening.

[keineahnung2345]

@wadmes This problem also occurs to me. And I found it calculates confidence as P(class|object), which doesn't take objectness into account. I revised some part of model.py(https://github.com/keineahnung2345/Mask_RCNN/commit/dfdfc7888ff7b74efd624082363a8eafff6ae043) and it calculate confidence as P(class|object)*P(object).
After this modification, it gives me a lower false positive rate.