I'm working using Mmdetection to train a Deformable DETR model using a custom COCO Dataset. Meaning a Custom Dataset using the COCO format of annotations. The dataset uses the same images as the COCO with different "toy" annotations for a "playground" experiment and the annotation file was created using the packages pycocotools and json exclusively.
I have made five variations of this playground dataset: 2 datasets with three classes (classes 1, 2, and 3), 1 dataset with six classes (classes 1 to 6) and 2 datasets with 7 classes (classes 1 to 7).
Now, after creating the dataset in mmdetection using mmdet.datasets.build_dataset, I used the following code to check if everything was OK:
from pycocotools.coco import COCO
from os import path as osp
from mmdet.datasets import build_dataset
cfg = start_config() # this is simply a function to startup the config file
ann_file = osp.join(cfg.data.train.data_root, cfg.data.train.ann_file)
coco = COCO(ann_file)
img_ids = coco.getImgIds()
ann_ids = coco.getAnnIds(imgIds=img_ids)
anns = coco.loadAnns(ids=ann_ids)
cats_counter = {}
for ann in anns:
if ann['category_id'] in cats_counter:
cats_counter[ann['category_id']]+=1
else:
cats_counter[ann['category_id']] = 1
print(cats_counter)
cats = {cat['id']:cat for cat in coco.loadCats(coco.getCatIds())}
for i in range(len(cats_counter)):
print("{} ({}) \t|\t{}".format(i, cats[i]['name'], cats_counter[i]))
ds = build_dataset(cfg.data.train)
print(ds)
For three of the datasets the amounts from the json file and from the constructed mmdet dataset are almost exactly equal. However, for one of the 3-classes dataset and for the 6-classes dataset, the results are incredibly different, where this code returns the following:
{3: 1843, 1: 659, 4: 1594, 2: 582, 0: 1421, 5: 498}
0 (1) | 1421
1 (2) | 659
2 (3) | 582
3 (4) | 1843
4 (5) | 1594
5 (6) | 498
loading annotations into memory...
Done (t=0.06s)
creating index...
index created!
CocoDataset Train dataset with number of images 1001, and instance counts:
+---------------+-------+---------------+-------+---------------+-------+---------------+-------+---------------+-------+
| category | count | category | count | category | count | category | count | category | count |
+---------------+-------+---------------+-------+---------------+-------+---------------+-------+---------------+-------+
| 0 [1] | 1421 | 1 [2] | 659 | 2 [3] | 581 | 3 [4] | 1843 | 4 [5] | 1594 |
| | | | | | | | | | |
| 5 [6] | 0 | -1 background | 45 | | | | | | |
+---------------+-------+---------------+-------+---------------+-------+---------------+-------+---------------+-------+
and
{1: 1420, 0: 4131, 2: 1046}
0 (1) | 4131
1 (2) | 1420
2 (3) | 1046
loading annotations into memory...
Done (t=0.06s)
creating index...
index created!
CocoDataset Train dataset with number of images 1001, and instance counts:
+----------+-------+------------+-------+----------+-------+---------------+-------+----------+-------+
| category | count | category | count | category | count | category | count | category | count |
+----------+-------+------------+-------+----------+-------+---------------+-------+----------+-------+
| | | | | | | | | | |
| 0 [1] | 1419 | 1 [2] | 0 | 2 [3] | 0 | -1 background | 443 | | |
+----------+-------+------------+-------+----------+-------+---------------+-------+----------+-------+
You can see that there is no "-1" id in the annotation json, and also some of the classes from the 3-classes dataset have 0 annotations, while the json clearly shows more than that. Has anyone encountered something similar using Mmdetection? What could be causing this problem?
There was a mismatch between the classes names in the annotation file and the classes names in the mmdetection config object. Correcting those solved the problem.