Issues about mAP
AXINLETTER opened this issue · 3 comments
Hello, the annotation information in the test set like:
{"license": 5, "file_name": "COCO_val2014_000000445200.jpg", "coco_url": "http://mscoco.org/images/445200", "height": 427, "width": 640, "date_captured": "2013-11-20 04:27:57", "flickr_url": "http://farm2.staticflickr.com/1147/5103880651_f3c1e2a721_z.jpg", "id": 445200}, and we use the model to get the detection result, like:
{"image_id": 445200, "category_id": 1, "score": 0.6065034866333008, "bbox": [380.34, 45.83, 86.32, 129.75], "COCO_category_id": 47},I can not understand which information means that this instance can complete that task in the anotation test set? I can get the score of an instance about being able to complete a task through the model but how to calculate the accuracy ? Thx~
I see that you have closed the issue, is it resolved?
I am so sorry that it took so long to think of this problem.
I saw it in the code
cocoEval = COCOeval(gtCOCO, dtCOCO, "bbox")
CocoEval. Params. CatIds = 1
CocoEval. The evaluate ()
CocoEval. The accumulate ()
CocoEval. The summarize ()
Then I got the accuracy rate of each task in the test set. I failed to figure out how to get the accuracy rate through the detection results. I hope you can answer, thank you!
For an object in the picture, the highest score of 14 tasks considered to be able to complete the task? Or should a threshold be set for the scores in the detection results, and objects larger than the threshold are considered to be able to complete the task? Or is it possible to extract the top n objects with the highest score as the prediction of the ability to complete the task? Or other evaluation methods?
If this is the first case, I can see that there are the same images in the test sets of different tasks, and an object in one image is also marked as able to complete multiple tasks. Obviously, the former is not feasible. In the second case, I don't see the specific score threshold. This makes me confused. I hope you can answer my question，Thx~