Using ImageBBox

Part 1 (2019)
#1

Hi,

I’m sorry probably a stupid question. I want to make a databunch for bounding box regression. I’m able to make the ImageBBox. However I can not find how to use these to get a labeled databunch.

I tried a lot of variations of:

dataDir = "./coco"
dataType = "val2017"
annFile = "{}/instances_{}.json".format(dataDir, dataType)
coco = COCOHumanBBoxDataset(annFile)
labels = {}
imgs = []
for id in coco.coco.getImgIds()[:12]:
    path, bbox = coco.get_one_pic_w_bb(id)

    imgs.append({"filename": path})
    labels[path.name] = bbox 
    name = Path(path).name

DATASET = pd.DataFrame.from_dict(imgs)

def get_y_func(pathStr):
    name = Path(pathStr).name
    return labels[name] # ImageBBox class

data = (
Data.ObjectItemList.from_df(
    df=DATASET, path=Path("/home/tako/devtools/furry-geras")
)
.random_split_by_pct()
.label_from_func(get_y_func)
.transform(Transform.get_transforms(), tfm_y=True, size=(224, 224))
.databunch()
)

An example using the ImageBBox class would be useful.

1 Like
#2

Hi, here’s an example for the pascal VOC dataset. You should be able to adapt it to COCO.

32
Screenshot 2019-02-16 at 11.14.32.png887×570 93.4 KB

1 Like
#3

Thanks this helps a lot. Also found this example using coco: