Sahi work on non axis oriented bounding box? Like yolo obb?

[karthikdatta98]

Does sahi work on non axis oriented bounding box models?

[whao159107]

Why is there no answer or example? Thank you.

[robmarkcole]

The answere is no, and trying to do so will fail with

prediction_result = [result.boxes.data for result in prediction_result]

AttributeError: 'NoneType' object has no attribute 'data'

This is because OBB are returned at result.obb. I also believe the sahi algo does not currently support OBB.
However the obb results do contain regular boxes too (xyxy), so it should be feasible to use those:

ultralytics.engine.results.OBB object with attributes:

cls: tensor([5.])
conf: tensor([0.2714])
data: tensor([[1.7813e+02, 1.3286e+02, 4.0502e+02, 3.2889e+02, 2.5792e+00, 2.7138e-01, 5.0000e+00]])
id: None
is_track: False
orig_shape: (1024, 1024)
shape: torch.Size([1, 7])
xywhr: tensor([[178.1333, 132.8626, 405.0189, 328.8881,   2.5792]])
xyxy: tensor([[ -80.8699, -114.2355,  437.1364,  379.9607]])
xyxyxyxy: tensor([[[ -80.8699,  101.7294],
         [  94.5004,  379.9607],
         [ 437.1364,  163.9958],
         [ 261.7661, -114.2355]]])
xyxyxyxyn: tensor([[[-0.0790,  0.0993],
         [ 0.0923,  0.3711],
         [ 0.4269,  0.1602],
         [ 0.2556, -0.1116]]])

[ZSBSB]

这样做应该是不行的

[robmarkcole]

This is supported in https://supervision.roboflow.com/latest/

[karthikdatta98]

Cant find where sahi is supported for obb.

[robmarkcole]

e.g.

 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), 
    # iou_threshold=0.3, 
    # overlap_filter_strategy='NON_MAX_SUPPRESSION'
)

detections = slicer(image)
detections = detections[detections.confidence > 0.5]
# detections = detections[detections.area > 1000]

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)

fig, ax = plt.subplots(figsize=(25, 20))
ax.imshow(annotated_image);
# ticks off
plt.axis('off');

[ZSBSB]

Thank you very much. Following the method you mentioned, I have completed the slice detection and also the detection of the rotating frame. Thank you so much! shaobo_zhang ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "obss/sahi" ***@***.***>; 发送时间: 2024年10月25日(星期五) 下午5:04 ***@***.***>; ***@***.******@***.***>; 主题: Re: [obss/sahi] Sahi work on non axis oriented bounding box? Like yolo obb? (Discussion #987) e.g. 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), # iou_threshold=0.3, # overlap_filter_strategy='NON_MAX_SUPPRESSION' ) detections = slicer(image) detections = detections[detections.confidence > 0.5] # detections = detections[detections.area > 1000] 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) fig, ax = plt.subplots(figsize=(25, 20)) ax.imshow(annotated_image); # ticks off plt.axis('off'); — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>