voc-->coco format conversion bug(?)

[jihyeonyi]

I've tried converting the voc data to coco format like belows.

$ datum import -i /mnt/hdd1/voc/VOCdevkit/VOC2012_detection/ -o prj-voc -f voc
$ datum filter -p prj-voc/ -e '/item[subset="train"]' -o prj-voc-train
$ datum export -p prj-voc-train/ --format coco -o voc-coco-train -- --tasks instances --save-images
$ datum project import -i voc-coco-train/ -f coco -o prj-voc-coco-train

For your information, "VOC2012_detection" is the same as VOC2012 except "ImageSets" folder.

ImageSets/
└── Main
    ├── train.txt
    └── val.txt

I've compared the distribution of bbox labels using validator, and I found that they're not same.

$ cd prj-voc-train/
$ datum validate detection
$ cd ../prj-voc-coco-train/
$ datum validate detection

from prj-voc-train

        "label_distribution": {
            "defined_labels": {
                "aeroplane": 470,
                "background": 0,
                "bicycle": 410,
                "bird": 592,
                "boat": 508,
                "bottle": 749,
                "bus": 317,
                "car": 1191,
                "cat": 609,
                "chair": 1457,
                "cow": 355,
                "diningtable": 373,
                "dog": 768,
                "foot": 0,
                "hand": 0,
                "head": 0,
                "horse": 377,
                "ignored": 0,
                "motorbike": 375,
                "person": 5019,
                "pottedplant": 557,
                "sheep": 509,
                "sofa": 399,
                "train": 327,
                "tvmonitor": 412
            },
            "undefined_labels": {}
        },
        "total_bbox_count": 15774

from prj-voc-coco-train

"label_distribution": {
            "defined_labels": {
                "aeroplane": 0,
                "background": 0,
                "bicycle": 0,
                "bird": 5717,
                "boat": 0,
                "bottle": 59,
                "bus": 0,
                "car": 0,
                "cat": 0,
                "chair": 4022,
                "cow": 0,
                "diningtable": 150,
                "dog": 0,
                "foot": 0,
                "hand": 0,
                "head": 0,
                "horse": 0,
                "ignored": 0,
                "motorbike": 0,
                "person": 5808,
                "pottedplant": 0,
                "sheep": 18,
                "sofa": 0,
                "train": 0,
                "tvmonitor": 0
            },
            "undefined_labels": {}
        },
        "total_bbox_count": 15774

In my opinion, the labels shouldn't be changed while converting the format.
I've also checked bbox size distribution(histogram), and it's also quite different.

[zhiltsov-max]

Hi, nice observation! Try to add --reindex 1 in coco export line to avoid confusing annotation ids in the output file.

[jihyeonyi]

Oh, it works! Thanks!