Tensorflow Object Detection API: how to find out false positives, false negatives, true positives?

Question

I am using Tensorflow Object Detection API to finetune a pretrained model from the model zoo for custom object detection. Once my model is converged I use eval_util.py with EvalConfig.metrics_set='open_images_V2_detection_metrics' to obtain the mAP (and class-specific AP s) which lets me measure the quality of my model.

But just mAP is not enough for my purposes. For better analysis, I want to know the exact breakdown of my model's results into false positives, false negatives and true positives. I wish to be able to see this breakdown in terms of actual test images - that is, I want to see my test images being physically divided into those three groups, automatically.

How can I do that?

I tried searching through Tensorflow's offical documentation and, to some extent, through the relevant python files on github, but I haven't found a way yet.

Answer 1

I think what you are looking for is a confusion matrix. Take a look at this link: Tensorflow Confusion Matrix

You can basically evaluate your predictions with this function.

Answer 2

We also meet this problem. Now we find some clues in object_detection/utils/metrics.py. Maybe you can have a try. Hope you can share your solutions!