No Change with Parameters updation

[usama-ayub-ch]

Hi IrDIE,
You did a great job for small object detection. I am using your implementation but when I change , slice_height, slice_width, overlap_height_ratio, overlap_width_ratio it does not change anything in the output same confusion matrix and results. Also when I added the confidence_threshold, in detection_model, I had not seen much changes. Can you please help me with that? Thank you

[IrDIE]

Hi! @usama-ayub-ch
1 - problem with changing hyperparameters ( slice_height, slice_width, overlap_height_ratio, overlap_width_ratio ) :
well, results gonna change if there are small objects that right in a boundary of sliced window. You can get results with overlap ratios (h and w) = 0, and check if anything changed.
Also: if sliced window inference make scence if you have many small objects, can you give example of images/labels you use?

2 - problem with changing confidence threshold:
can you describe how exactly you are changing it?
There are argument conf in compile_predictor() function that change conf threshold

• what ultralitycs version you use?

[usama-ayub-ch]

Hi,
Thank you for your response.
I have 2 classes but multiple small objects and I have satellite Images.
I am using Ultralytics 8.3 and I tried to add confidence threshold in the function, AutoDetectionModel.from_pretrained().

Thanks

[usama-ayub-ch]

Hi! @usama-ayub-ch 1 - problem with changing hyperparameters ( slice_height, slice_width, overlap_height_ratio, overlap_width_ratio ) : well, results gonna change if there are small objects that right in a boundary of sliced window. You can get results with overlap ratios (h and w) = 0, and check if anything changed. Also: if sliced window inference make scence if you have many small objects, can you give example of images/labels you use?

2 - problem with changing confidence threshold: can you describe how exactly you are changing it? There are argument conf in compile_predictor() function that change conf threshold

• what ultralitycs version you use?

How can I change 'conf threshold' in validation function.
detection_model = AutoDetectionModel.from_pretrained(
model_type='yolov8',
model_path=yolov8_model_path,
confidence_threshold=0.3,
device="cuda:0" # or "cpu"
)

like this. Because its detecting images with conf=0.3. I want the images with conf >4