Mask_rcnn: show low detection score results

[patrick-llgc]

First of all, thanks for making this code public and thanks to everyone who has been contributing to it.

I am running into some issues and I wonder if some one here can enlighten me.

I am running anomaly detection on some images, and I only have one class (plus the background). All returned results seems to have detection scores higher than 0.5, even though I set DETECTION_MIN_CONFIDENCE to 0. How could I show detection with scores smaller than 0.5? Or if you are also doing something similar, are you able to showcase detection results with low scores?

Thanks!

[patrick-llgc]

OK I've found the cause of my problem.

If there are only two classes (one object class + background), then only detection score > 0.5 will be returned, as bboxes with detection score <0.5 will be classified as background. The detection layer removes the background bboxes and thus it only returns bboxes with detection score > 0.5.

The logic is implemented in model.refine_detections_graph.

To return everything with score < 0.5, change the first line

# Class IDs per ROI
class_ids = tf.argmax(probs, axis=1, output_type=tf.int32)

to

# Class IDs per ROI
if config.NUM_CLASSES == 2:
    class_ids = tf.ones_like(probs[:, 0], dtype=tf.int32)
else:
    class_ids = tf.argmax(probs, axis=1, output_type=tf.int32)

Basically sets all class_ids to 1 and now it returns bboxes with small detection scores. Now you can filter it with config. DETECTION_MIN_CONFIDENCE as you need.

[Dgs29]

@patrick-llgc . I know it has been a while since you posted, but I have some doubts. I also have two classes(1 object class + background). Why does the model predict only objects with scores> 90%? Where should I change if I want to find results that are lower than that?