[BUG] Bbox coordinate mismatch in hub visualization

🐛🐛 Bug Report

⚗️ Current Behavior

In Hub documentation, BBox is defined as tensorflow’s style of normalized coordinate in [ymin, xmin, ymax, xmax]. But when I want to visualize it in app.activeloop.ai, the bbox was very small at the top-left corner.

Input Code

h, w, c = img.shape
box = np.array([y1/h, x1/w, y2/h, x2/w])

Expected behavior/code The bbox should be in the same encoding type as in the dataset and hub visualizer

⚙️ Environment

• Python version(s): 3.6.9
• Google Colab

🧰 Possible Solution

After several tests, PASCAL VOC’s style of unnormalized [xmin, ymin, xmax, ymax] is the solution to correctly visualize the bbox.

box = np.array([x1, y1, x2, y2])

🖼 Additional context/Screenshots

A. Using Hub’s style [ymin, xmin, ymax, xmax] small red dot at the top corner

B. Using Pascal’s stye [xmin, ymin, xmax, ymax] with segmentation too

[AbhinavTuli]

@Diveafall can you take a look at this?

[haritsahm]

@haritsahm Can you see the bounding boxes now?

Yes I can!

@AbhinavTuli @haritsahm is the default bbox format normalized or unnormalized now?

You can use both but you have to choose the datatype in the visualizer app, but the official documentation uses normalized [xmin, ymin, xmax, ymax]

You can use the **unnormalized [xmin, ymin, xmax, ymax] ** and set the bbox datatype with Basic [xmin, ymin, xmax, ymax] (the first option)

You can also use the Tensorflow’s version of **normalized [ymin, xmin, ymax, xmax] ** and set the bbox datatype with the Tensorflow Bounding Box (third option).

The second option is unusual for me. It only works with unnormalized bbox with a format of:

top_left_y = ymin-box_height# shift the coordinate
bottom_right_y = ymax-box_height# shift the coordinate

bbox = [top_left_y , xmin, bottom_right_y , xmax] # unnormalized

I’m not sure why I have to shift y coordinate. Maybe @Diveafall or @AbhinavTuli can check it out.

I think another good option for a bounding box encoding are:

• normalized/unnormalized [xmin, ymin, box_width, box_height]
• YOLO’s style normalized [xmin, ymin, box_width, box_height] relative to width and height of image

I’m sorry for closing this issue without testing it because I think that is enough for me. I should have tested it first before closing this issue, Feel free to open the issue again!