Maskrcnn-benchmark: Only bounding box annotation

[DecentMakeover]

❓ Questions and Help

Hi

Thanks a lot for sharing your work.
for my dataset, i have bounding box annotations but i don't have the masks, is there any chance i could still train the network by box co-ordinates.

Thanks

[LeviViana]

Yes you can. You just need to pick an architecture that only trains on bounding boxes, like all e2e_faster_rcnn. Though, be sure you have labelled you data in coco-style. Maybe this thread can give you a hint on that matter.

[DecentMakeover]

Hi

because my data is mostly rectangular , i wrote a function to get masks from my bounding boxes.

so hopefully i should be able to train mask-rcnn!
Thanks for your help.

[DecentMakeover]

Hi @fmassa @LeviViana

Few questions about the annotations format,

i am going to use the detection format, so the required fields are

1. area 2. bbox 3.category_id 4.image_id 5.iscrowd 6.segmentation

Im guessing this is all i need to train my dataset, kindly correct me if i'm wrong.

--The bbox field it takes in four points- im guessing this is x1,y1,x2,y2- atleast my points are in this format.

--in the segmentation field it takes in rle or polygon - Can i pass in it a list of list of eight points, it will look like this -[[[0,45],[336,45],[336,270],[0,270]]] , will this work?

-- Area, i'm guessing this is the area of the mask, am i just supposed to sum all the positive pixels(white pixels)?, will that work?

Clarifications on this will be helpful,

Thanks in advance.

[fmassa]

@DecentMakeover the data is expected to be in COCO format.
Thus the boxes should be in x0, y0, w, h, but this can be easily modified https://github.com/facebookresearch/maskrcnn-benchmark/blob/90080e60cc4a9252d019f475af044c9d4119b09b/maskrcnn_benchmark/data/datasets/coco.py#L75

If you don't have segmentation data, just remove https://github.com/facebookresearch/maskrcnn-benchmark/blob/90080e60cc4a9252d019f475af044c9d4119b09b/maskrcnn_benchmark/data/datasets/coco.py#L82-L84 and you'll be good to go

For the area, it's the area of the polygon. If you don't have polygons, just use the are of the bounding box.

[AdanMora]

Hi, you can easily generate the coco data using only the bounding box or segmentation polygon, I did it manually based on this repo

[DecentMakeover]

@fmassa Thanks a lot for the response.

@AdanMora , okay Thanks, ill look into it.