simulate_sahi.py

import cv2
import numpy as np
import torch
import time
import argparse
import os
from ultralytics import YOLO
from download_model import is_lfs_pointer, download_file

def slice_image(image, slice_size=640, overlap=0.2):
    height, width, _ = image.shape
    step = int(slice_size * (1 - overlap))
    
    slices = []
    for y in range(0, height, step):
        for x in range(0, width, step):
            y_end = min(y + slice_size, height)
            x_end = min(x + slice_size, width)
            y_start = max(0, y_end - slice_size)
            x_start = max(0, x_end - slice_size)
            
            slice_img = image[y_start:y_end, x_start:x_end]
            slices.append({
                'image': slice_img,
                'x_offset': x_start,
                'y_offset': y_start
            })
            if x_end == width:
                break
        if y_end == height:
            break
    return slices

def run_sliced_inference(model, image, slice_size=640, overlap=0.2, conf=0.25, iou_threshold=0.45):
    slices = slice_image(image, slice_size, overlap)
    all_boxes = []
    all_confs = []
    all_clss = []
    
    t_start = time.time()
    for s in slices:
        results = model.predict(s['image'], conf=conf, verbose=False)
        for result in results:
            boxes = result.boxes
            for box in boxes:
                xyxy = box.xyxy[0].cpu().numpy()
                global_xyxy = [
                    xyxy[0] + s['x_offset'],
                    xyxy[1] + s['y_offset'],
                    xyxy[2] + s['x_offset'],
                    xyxy[3] + s['y_offset']
                ]
                all_boxes.append(global_xyxy)
                all_confs.append(float(box.conf[0]))
                all_clss.append(int(box.cls[0]))
    t_inference = time.time() - t_start
                
    if len(all_boxes) == 0:
        return [], t_inference
        
    boxes_tensor = torch.tensor(all_boxes)
    confs_tensor = torch.tensor(all_confs)
    
    # Standard NMS (Non-Maximum Suppression) to merge detections across slices
    from torchvision.ops import nms
    keep_indices = nms(boxes_tensor, confs_tensor, iou_threshold=iou_threshold)
    
    filtered_results = []
    for idx in keep_indices:
        idx = int(idx)
        filtered_results.append({
            'box': all_boxes[idx],
            'conf': all_confs[idx],
            'class': all_clss[idx]
        })
    return filtered_results, t_inference

def main():
    parser = argparse.ArgumentParser(description="Simulate Standard vs SAHI Sliced Inference")
    parser.add_argument("--image", type=str, default="sample_aerial_street.jpg", help="Path to input image")
    parser.add_argument("--model", type=str, default="aerialEye.pt", help="Path to YOLO model")
    parser.add_argument("--conf", type=float, default=0.25, help="Confidence threshold")
    parser.add_argument("--slice-size", type=int, default=640, help="SAHI slice dimension")
    parser.add_argument("--overlap", type=float, default=0.2, help="Overlap ratio between slices")
    
    args = parser.parse_args()
    
    # Check/download files if missing
    for file_path in [args.model, args.image]:
        if file_path in ["aerialEye.pt", "best.pt", "sample_aerial_street.jpg", "sample_drone_roundabout.jpg", "bus.jpg"] and (not os.path.exists(file_path) or is_lfs_pointer(file_path)):
            print(f"File '{file_path}' is missing or is an LFS pointer. Downloading...")
            download_file(file_path)
            
    if not os.path.exists(args.image):
        print(f"Error: Image '{args.image}' not found.")
        return
        
    print(f"Loading YOLO model '{args.model}'...")
    model = YOLO(args.model)
    
    img = cv2.imread(args.image)
    if img is None:
        print(f"Error: Could not read image '{args.image}'.")
        return
        
    h, w, _ = img.shape
    print(f"Loaded image: {args.image} ({w}x{h})")
    
    # 1. Run Standard Inference
    print("\n--- Running Standard Inference ---")
    t_start = time.time()
    std_results = model.predict(img, conf=args.conf, verbose=False)
    std_time = time.time() - t_start
    std_boxes = std_results[0].boxes
    print(f"Standard Detections: {len(std_boxes)}")
    print(f"Standard Latency: {std_time:.4f} seconds")
    
    # Save standard image
    std_img = std_results[0].plot()
    cv2.imwrite("result_standard.jpg", std_img)
    print("Saved standard inference output to 'result_standard.jpg'")
    
    # 2. Run Sliced Inference
    print("\n--- Running SAHI Sliced Inference Simulation ---")
    slices = slice_image(img, args.slice_size, args.overlap)
    print(f"Generated {len(slices)} slices of size {args.slice_size}x{args.slice_size} with {args.overlap*100}% overlap.")
    
    sahi_results, sahi_time = run_sliced_inference(model, img, args.slice_size, args.overlap, args.conf)
    print(f"SAHI Detections (after NMS): {len(sahi_results)}")
    print(f"SAHI Total Latency: {sahi_time:.4f} seconds")
    
    # Render SAHI image
    sahi_img = img.copy()
    for res in sahi_results:
        box = res['box']
        cls_id = res['class']
        conf_val = res['conf']
        # Draw bounding box
        cv2.rectangle(sahi_img, (int(box[0]), int(box[1])), (int(box[2]), int(box[3])), (0, 255, 0), 2)
        # Put label
        label = f"{model.names[cls_id]} {conf_val:.2f}"
        cv2.putText(sahi_img, label, (int(box[0]), int(box[1]) - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
        
    cv2.imwrite("result_sahi.jpg", sahi_img)
    print("Saved SAHI inference output to 'result_sahi.jpg'")
    
    # 3. Print Comparison
    print("\n--- Summary Comparison ---")
    print(f"| Metric | Standard | SAHI (Sliced) | Difference |")
    print(f"|---|---|---|---|")
    print(f"| Objects Detected | {len(std_boxes)} | {len(sahi_results)} | {len(sahi_results) - len(std_boxes):+d} |")
    print(f"| Inference Latency | {std_time:.3f}s | {sahi_time:.3f}s | {sahi_time - std_time:+.3f}s |")
    
if __name__ == "__main__":
    main()