Clarification of obb behaviour with InferenceSlicer

[robmarkcole]

Search before asking

• I have searched the Supervision issues and found no similar feature requests.

Question

Based on reviewing source code, I understand that with an obb model, the regular box enclosing the obb box is used with SAHI. As a result, merged regular boxes, and not obb boxes are returned. I wanted to confirm this, and suggest a note is added in the docs.

I also want to know if this could be responsible for the odd predictions below - perhaps the enclosing boxes are not merged correctly:

Additional

No response

[SkalskiP]

Hi, @robmarkcole 👋🏻 Lots to unpack here.

• OBB support in InferenceSlicer has not been officially released yet, but we're working on it. (We should perhaps explicitly state that only object detection and instance segmentation are currently supported.)
• A test version of OBB in InferenceSlicer is available in the develop branch. You can read more about it here and here. To use it, you would need to install supervision like this: pip install git+https://github.com/roboflow/supervision.git.
• If you want to plot OBB, use OrientedBoxAnnotator instead of BoundingBoxAnnotator (deprecated) or BoxAnnotator.
• Regarding the quality of the result you received, it might be related to the second issue you opened - incorrect channel order. What slice_wh value are you using? The default is (320, 320), I'd try to increase it to (640, 640).

[robmarkcole]

Thanks @SkalskiP these results are much better!

image_path = "my.png"
image = np.array(Image.open(image_path))

def callback(image_slice: np.ndarray) -> sv.Detections:
    result = model(image_slice)[0]
    return sv.Detections.from_ultralytics(result)

slicer = sv.InferenceSlicer(callback = callback, slice_wh=(768, 768))
detections = slicer(image)
detections = detections[detections.class_id == 1] # ships id in DOTA

oriented_box_annotator = sv.OrientedBoxAnnotator()
label_annotator = sv.LabelAnnotator()

annotated_image = oriented_box_annotator.annotate(scene=image, detections=detections)
annotated_image = label_annotator.annotate(scene=annotated_image, detections=detections)

[SkalskiP]

@robmarkcole that is true; this looks a lot better! There is still room for improvement. Especially where it comes to post-processing and merging double detections. We don't have that component yet for OBB.

Can I close the issue?

[050603]

`import cv2
import numpy as np
import supervision as sv
from ultralytics import YOLO
from PIL import Image

model = YOLO("/root/autodl-tmp/code/ultralytics/runs/obb/train/weights/best.pt")

image_path = "/root/autodl-tmp/DOTA/images/test/P1068.png"
image = np.array(Image.open(image_path))

def callback(image_slice: np.ndarray) -> sv.Detections:
result = model(image_slice)[0]
return sv.Detections.from_ultralytics(result)

slicer = sv.InferenceSlicer(callback = callback)
detections = slicer(image)

oriented_box_annotator = sv.OrientedBoxAnnotator()
label_annotator = sv.LabelAnnotator()

annotated_image = oriented_box_annotator.annotate(scene=image, detections=detections)
annotated_image = label_annotator.annotate(scene=annotated_image, detections=detections)

annotated_pil_image = Image.fromarray(annotated_image)
annotated_pil_image.show()`

@SkalskiP Why doesn't the target box display at the correct scale when drawing with OrientedBoxAnnotator?

[LinasKo]

I believe we have a fix in the develop branch.

Did you try this?
pip install git+https://github.com/roboflow/supervision.git@develop

[050603]

I believe we have a fix in the develop branch.

Did you try this? pip install git+https://github.com/roboflow/supervision.git@develop

@LinasKo Yes, that's how I do it.

[050603]

@LinasKo I checked the saved CSV file and found that the coordinates were all relative to the pixel position of the slice (up to 640), without considering the position of the slice itself.

[SkalskiP]

Hi @050603 👋🏻 is there any chance you could provide us with your model and image?