Access to probability distribution of detection

[leoandeol]

Currently after prediction on a dataset, the resulting output contains for each image, a dictionary "pred_instances", which has three members itself: labels, scores, and bboxes. I would like to get for each predicted box, both the softmax distribution (not just the argmax), as well as additional scores (for instance the objectness score from the YOLO models).

How can I obtain such information?

[Haleski47]

Currently after prediction on a dataset, the resulting output contains for each image, a dictionary "pred_instances", which has three members itself: labels, scores, and bboxes. I would like to get for each predicted box, both the softmax distribution (not just the argmax), as well as additional scores (for instance the objectness score from the YOLO models).

How can I obtain such information?

Hello Haian huang. I meet the same problem. I want to get the distribution but have no solution. May i know how you solve the problem?

[leoandeol]

I solved it by not using mmdet and instead torch hub models.