jskinn / rvchallenge-starter-kit Goto Github PK

I'm having some trouble interpreting the json ground truth detections for the validation data set, as the json file looks a bit different from what I was expecting.

For example, the first few entries of validation_data/ground_truth/000000/labels.json look like:

{"005171": {
            "_metadata": {
                        "mask_name": "3813.masks.png"
                        },
            "000046":   {
                        "mask_id": 1, "bounding_box": [219, 42, 223, 48], "class": "bottle", "num_pixels": 19
                        },
            "000131":  {"mask_id": 2, "bounding_box": [482, 0, 564, 57], "class": "potted plant", "num_pixels": 2104
                        }
            },
"004500": {
            "000127": {
                        "mask_id": 1, "bounding_box": [612, 58, 639, 369], "class": "oven", "num_pixels": 8338
                      },
          "_metadata": {
                      "mask_name": "3391.masks.png"
                      }
          },

How do the top level keys (like "005171" or "004500" above) map to specific entries in the dataset?

I guess I can extract the corresponding image name from the _metadata: 'XXXX.masks.png' entry but this doesn't seem to match with what Submission Format or Validation Data indicate about the format in the readme, so I am slightly confused.

Thanks!

This is barely a bug, but I found when applying a coco detector to this challenge that your class list uses 'television' while the COCO category names use 'tv'. It's pretty easy to handle on our end, but I figured I'd open an issue for this in case you wanted them to exactly match with COCO.

Thanks for hosting this challenge!