🚗 YOLOv11 + SAHI — Improved Small Object Detection

Small object detection is one of the hardest open problems in computer vision. In a compact traffic scene, cars closest to the camera are large and easy to detect — but vehicles further down the road shrink to a handful of pixels and are frequently missed entirely by standard detectors.

This project trains a YOLOv11 model for vehicle detection and augments it with SAHI (Slicing Aided Hyper Inference) — a technique that slices high-resolution images into overlapping tiles, runs inference on each tile, and merges the results — to recover detections on small, distant, and dense objects without retraining the model.

📽️ Originally presented as an academic project at Rajkiya Engineering College, Kannauj (AKTU), guided by Mr. Shashank Yadav.

• High-resolution images get resized before inference, and small objects lose critical visual detail in the process.
• Tiny/distant objects occupy very few pixels, so standard YOLO struggles to detect them.
• In dense traffic scenes, vehicles further from the camera appear progressively smaller — leading to missed detections for anything but the foreground.
• Retraining a bigger/higher-resolution model is expensive; the goal here is to improve detection at inference time only.

• Study small object detection techniques in deep learning.
• Train a YOLOv11 object detection model on a custom vehicle dataset.
• Run standard YOLOv11 inference as a baseline.
• Apply SAHI (slicing + merged inference) on top of the trained model.
• Compare YOLOv11 vs YOLOv11 + SAHI on detection count, small-object recall, and accuracy trade-offs.

• Prepared and exported in YOLO/COCO format via Roboflow.
• Contains annotated traffic scenes with classes such as Car, Bus, Lorry/Truck, Motorcycle.
• Split into training, validation, and test sets.
• Downloaded directly into Google Drive for use inside the Colab training pipeline.

1. Dataset preparation — annotate and export dataset using Roboflow.
2. Training — fine-tune YOLOv11 (yolo11m.pt) on the custom dataset using the Ultralytics CLI.
3. Baseline inference — run standard YOLOv11 inference on test images.
4. Image slicing — use SAHI to slice each high-resolution image into overlapping tiles.
5. Sliced inference — run YOLOv11 inference independently on every tile.
6. Merge predictions — recombine all tile-level detections back into full-image coordinates and apply Non-Maximum Suppression (NMS) to remove duplicate boxes.

# Train YOLOv11 on custom dataset
yolo train model=yolo11m.pt data="data.yaml" epochs=50 imgsz=640

# Validate trained model
yolo val model="runs/detect/train/weights/best.pt" data="data.yaml"

from sahi.auto_model import AutoDetectionModel
from sahi.predict import get_sliced_prediction

detection_model = AutoDetectionModel.from_pretrained(
    model_type="yolov11",
    model_path="runs/detect/train/weights/best.pt",
    confidence_threshold=0.25,
    image_size=640
)

result = get_sliced_prediction(
    image_path,
    detection_model,
    slice_height=512,
    slice_width=512,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2
)

Inference: SAHI improves detection of small and dense objects — a +37% increase in total detected objects in the same scene, largely from vehicles further down the road that the baseline model missed.

• ✅ YOLOv11 alone detects foreground/large objects effectively.
• ✅ YOLOv11 + SAHI recovers significantly more small, distant, and occluded vehicles.
• ⚠️ Trade-off: slicing can introduce a drop in per-detection confidence/accuracy on some tiles (partial objects at slice boundaries, duplicate/fragmented boxes before NMS cleanup) — this is an active limitation, see below.
• ✅ Works entirely at inference time — no retraining required.

• Accuracy trade-off with SAHI: while SAHI increases the number of small objects detected, it can reduce per-box confidence/precision in some cases — likely due to objects being split across slice boundaries or duplicate detections not being fully resolved by NMS. Tuning overlap_height_ratio, overlap_width_ratio, and the NMS IoU threshold is an open area for improvement.
• Slicing increases inference time proportionally to the number of tiles — a real-time/latency trade-off for the accuracy gain.
• Very small/heavily occluded objects at the tile edges can still be missed or double-counted pre-NMS.

• Useful for identifying tiny objects that create outsized problems in real-world scenarios.
• Missing small objects can lead to serious safety and operational risks (e.g., missed pedestrians/vehicles in traffic monitoring).
• Applicable to surveillance, traffic monitoring, autonomous systems, drone imagery, and satellite imagery.
• Can be extended to medical imaging for detecting small but clinically significant anomalies.
• Further tuning of slice size/overlap and NMS thresholds to close the accuracy gap introduced by slicing.

• Small object detection remains a core challenge in computer vision.
• Standard YOLOv11 struggles with tiny/distant objects in dense scenes.
• SAHI meaningfully improves small-object recall through slicing-based inference.
• This improvement is achieved without retraining the underlying model — making it a practical, low-cost enhancement for any existing YOLO deployment.

# Clone the repo
git clone <your-repo-url>
cd <repo-name>

# Install dependencies
pip install ultralytics sahi

# Train
yolo train model=yolo11m.pt data="data.yaml" epochs=50 imgsz=640

# Run baseline inference
yolo predict model="runs/detect/train/weights/best.pt" source="path/to/image.jpg"

# Run SAHI-enhanced inference — see notebook for full script

The full, reproducible pipeline (dataset setup → training → baseline inference → SAHI-sliced inference → comparison) is available in the notebook: YOLO11_and_sahi_for_improved_small_object_detection.ipynb

• Author: Ayush Kumar Singh
• Guide: Mr. Shashank Yadav
• Rajkiya Engineering College, Kannauj — Dr. A.P.J. Abdul Kalam Technical University