Mask R-CNN on NYUv2

Mask R-CNN on NYUv2

This repository mainly contains information from the execution of the Mask R-CNN network [1] on images from the NYUv2 dataset [2] as well as additional metadata. It was created for analyzing the output of Mask R-CNN and post-processing it using contextual information for improving its performance. This work has been carried out by Dr. Jose-Raul Ruiz-Sarmiento (MAPIR group, University of Málaga) and Dr. Shuda Li (AVG group, University of Oxford) in the scope of the European project MoveCare: Multiple-actOrs Virtual Empathic CARgiver for the Elder (Ref: 732158).

Concretely, this repository includes:

- metadata:
    + coco_nyu_mapping.txt: Mapping between the categories in COCO dataset and those in NYUv2.
    + coco_object_categories.txt: Object categories considered in COCO dataset.
    + nyu_object_categories.txt: Object categories used in NYUv2 dataset.
    + nyu_scene_categories.txt: Scene categories considered in NYUv2.
    + objects_and_categories_in_images.txt: For each image in NYUv2, the categories of the appearing objects.

- nyu_content:
    + masks_in_X (Where X is the image index)
        - Y.png: Where Y is the object index in the image, represents the binary mask of that object.
        - pixels_labelled.png: Binary mask indicating the labelled pixels in image X.
    + bboxesX.txt: Where X is the image index, includes the ground truth bounding boxes of the objects in it. Format is: min_x min_y max_x max_y.

- preds:
    + X: Where X is the image index.
        - Y.png: Where Y is the object index in the image, as detected by Mask R-CNN. Binary image containing the mask of such detected object.
    + X.txt: Where X is the image index. File containing the objects detected by Mask R-CNN, including: idx class score min_x min_y max_x max_y masks_file, being min_x min_y max_x and max_y bounding box information, while masks_file refers to X/Y.png as described above.
    + result_X.png: Where X is the image index. Image showing the detections with a socre higher than 0.3.
    + gt_iou_X: Where X is the image index.
        - Y: Where Y is the index of the detected object.
            + Z.png Where Z is the index of the object in the ground truth. Image showing the masks of both objects, Y and Z, for visually checking their overlapping.
        - Y.txt: Where Y is the index of the detected object. File containing:
            + The intersection ratio of the object mask Y with the labelled part of the image.
            + The IoU value for the mask of object Y and those of ground truth objects.

References:

[1] He, Kaiming, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. "Mask r-cnn." In Proceedings of the IEEE international conference on computer vision, pp. 2961-2969. 2017.
[2] Silberman, Nathan, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. "Indoor segmentation and support inference from rgbd images." In European Conference on Computer Vision, pp. 746-760. Springer, Berlin, Heidelberg, 2012.