SAHI inference with YOLOv8 💫 boost your small object detection performance

What is SAHI (Slicing Aided Hyper Inference) - https://docs.ultralytics.com/guides/sahi-tiled-inference/

What in this repo:

This repo contain not only SAHI inference implementation but also evaluation of results with mAp50, ... (standart metrics).

If imgsz = None in main() function :

• prediction will be done with original imagesize
• validation will be done with original imagesize ( force batch size to 1 )

You will understand if SAHI inference help in your specific case

SAHI	No SAHI
more cars far away detected	standart detections

Run output:

• SAHI inference + EVALUATION of results with basic yolo8 metrics
  • output example with basic validation on 2 images:

          val: Scanning C:\Users\irady\GitHub\YOLO8_SAHI\yolo_dataset\labels.cache.
                     Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 
                       all          2         11      0.987      0.545       0.57      0.455
                    person          2         11      0.987      0.545       0.57      0.455
    Speed: 1.5ms preprocess, 329.6ms inference, 0.0ms loss, 4.5ms postprocess per image
    
        ```

• output example with SAHI validation on 2 images:

  • val: Scanning C:\Users\irady\GitHub\YOLO8_SAHI\yolo_dataset\labels.cache.
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 
    Performing prediction on 9 number of slices.
    Performing prediction on 9 number of slices.
             Class     Images  Instances      Box(P          R      mAP50  mAP50-95):
               all          2         11          1      0.545      0.773      0.628
            person          2         11          1      0.545      0.773      0.628
    Speed: 7.5ms preprocess, 0.0ms inference, 0.0ms loss, 0.0ms postprocess per image
      

• and also check sahi/ folder - there all validation plots will be saved

How to use:

• git clone
• in utils.get_category_mapping() change returned dictionary for your classes
• in main() change paths for your .pt and .yaml, and set desired input imgsz, source for inference etc
• VALIDATION : in main() run run_sahi_validation() or run_basic_validation()
• INFERENCE : in main() run run_sahi_prediction() or run_basic_prediction()
  • also you can update size of sliding window in head of utils.sahi_predict() :

        VERBOSE_SAHI = 2
        SLICE_H = 640
        SLICE_W = 640
        OVERLAP_HEIGHT_RATIO = 0.2
        OVERLAP_WIDTH_RATIO = 0.2```
    
    

What you need:

• your .pt model file (or any format that ultralytics support - like .onnx or .engine)
• validation dataset in yolo standart format
• .yaml file for dataset

  • #file sahi_data.yaml
    path: ../YOLO8_SAHI/yolo_dataset/ # dataset root dir
    
    train: images # train images (relative to 'path') 128 images
    val: images # val images (relative to 'path') 128 images
    
    # Classes
    names:
      0: people