@spec postprocess(
  model :: YOLO.Model.t(),
  model_output :: Nx.Tensor.t(),
  scaling_config :: YOLO.FrameScalers.ScalingConfig.t(),
  opts :: Keyword.t()
) :: [[float()]]

Post-processes the model's raw output to produce a filtered list of detected objects.

The raw output tensor has shape {1, 84, 8400} where:

• First dimension (1) is the batch size
• Second dimension (84) contains 4 bbox coordinates + 80 class probabilities per detection
• Third dimension (8400) is the number of candidate detections

The processing steps are:

1. Reshapes and transposes output to {8400, 84} format
2. Applies non-maximum suppression (NMS) to filter overlapping detections
3. Scales bounding boxes back to original image dimensions

Arguments

• model - YOLO.Model struct containing model metadata
• model_output - Raw output tensor from model inference {1, 84, 8400}
• scaling_config - Scaling configuration from preprocessing step
• opts - Keyword list of options:
  • prob_threshold - Minimum probability threshold for detections
  • iou_threshold - IoU threshold for non-maximum suppression
  • nms_fun - Optional custom NMS function (defaults to YOLO.NMS.run/3)

Returns

List of detections where each detection is a list [cx, cy, w, h, prob, class_idx]:

• cx, cy - Center coordinates of bounding box
• w, h - Width and height of bounding box
• prob - Detection probability
• class_idx - Class index (0-79 corresponding to COCO classes)

@spec preprocess(YOLO.Model.t(), term(), Keyword.t()) ::
  {Nx.Tensor.t(), YOLO.FrameScalers.ScalingConfig}

Preprocesses an input image to match the model's required format.

The preprocessing steps are:

1. Scales and pads the image to match model input dimensions while preserving aspect ratio
2. Converts RGB to BGR color format (Ultralytics models expect BGR input)
3. Normalizes pixel values to [0,1] range by dividing by 255
4. Transposes dimensions to match model's expected format
5. Adds batch dimension

Arguments

• model - YOLO.Model struct containing model metadata
• image - Input image in implementation's native format
• options - Keyword list of options including :frame_scaler module

Returns

• {input_tensor, scaling_config} tuple where:
  • input_tensor has shape {1, 3, height, width}
  • scaling_config contains scaling/padding info for postprocessing