Maskrcnn-benchmark: Inference pth results and prediction pth files

[Idolized22]

❓ Questions and Help

Hello,
I have few question about testing self created dataset,
I have used the e2e_faster_rcnn_X_101_32x8d_FPN_1x.ymal model .

1. what and how does the results.pth and prediction.pth used for ?
2. does one of them holds the mask predictions ?
   3.is there a way to get a json polygons prediction file for the test set ?
3. in order to visualize the results do I need to modify the predictor.py categories ?
   does some more modifications are required ?

Thanks for your help in advance !

[fmassa]

Hi,

Here is a thread that is related to your questions (specifically point 3) https://github.com/facebookresearch/maskrcnn-benchmark/issues/327

About the other points:
1 - Those files are intermediate results of the inference, and they contain different information. coco_results.pth is just a struct containing the mIoU, while prediction.pth holds the outputs of the model for every image in the dataset
2 - yes, the masks are stored in prediction.pth, but before being pasted in the image (so you need to use the Masker to paste the masks into the image space, see https://github.com/facebookresearch/maskrcnn-benchmark/blob/de42d895ad7c2cee72259207fa9f2a7580c72cd0/maskrcnn_benchmark/data/datasets/evaluation/coco/coco_eval.py#L108 and https://github.com/facebookresearch/maskrcnn-benchmark/blob/de42d895ad7c2cee72259207fa9f2a7580c72cd0/maskrcnn_benchmark/data/datasets/evaluation/coco/coco_eval.py#L122-L125
3 - the rle encodings are available in the segm.json file, you could try seeing if you can get the polygons out of it

I'm closing the issue, but let me know if you have further questions