COCO Matlab API repurposed to evaluate zero-shot segmentation performance on unseen categories i.e. the 60 COCO categories that are not in common with the 20 categories of the PASCAL dataset.
Includes 3 additions:
-
sj_UnseenCat_SaveImages.m
- Saves coco-images that contain objects all which must belong to categories not in the 20 PascalVOC categories.
-
sj_unseen_eval_1.m
- Stores ground-truth and predictions (for all the images saved above) in COCO JSON format as separate .mat files % (All the predicted categories get single label - 81 (unseen cat label))
-
sj_unseen_eval_2.m
- Estimates MAP scores for zero-shot segmentation performance on the images of unseen categories saved in step-1 and predicted using Straight to shapes approach implemented as this.