Research Area 2 · Object Detection & Instance Segmentation

Research 2: YOLO Object Detection & SAM 3 Segmentation

Converting classroom images into structured object detections, then refining each bounding box into a pixel-level segmentation mask for tracking, inventory, and scene comparison.

Individual research by Shriya Paladugu

A classroom is a constantly changing environment where people and equipment enter, leave, or move locations. A camera records pixels, but raw images alone do not tell the system what objects are present or where each object appears. The project therefore needs a computer-vision pipeline that identifies classroom objects, records their confidence scores and image locations, and separates them from the surrounding background.

How accurately and efficiently can a locally executed YOLO and SAM 3 pipeline detect, classify, locate, and segment important classroom objects under changing lighting, viewing angles, distance, and partial occlusion while remaining within the target 10-second processing cycle?

flowchart TD
    CAP["Camera capture\nTimestamped frame"] --> PRE["OpenCV / FFmpeg\nResize + preprocessing"]
    PRE --> YOLO["YOLO\nClass + confidence + box"]
    YOLO --> SAM["SAM 3\nPixel-level mask"]
    SAM --> OUT["Structured observation\nDetection + mask + timestamp"]
    OUT --> TRACK["Tracking / inventory /\nscene comparison"]
Why both models? YOLO identifies what an object is and provides its bounding box. SAM 3 uses that box as a prompt to trace the object's more precise pixel boundary. The segmentation mask can then support spatial comparison and object-state tracking.

The initial prototype will evaluate six common classroom classes that can support people counting and basic equipment inventory.

Person

Supports occupancy counting without facial recognition.

Laptop

Represents portable classroom hardware.

Backpack

Tests detection of personal items and partial occlusion.

Chair

Provides a frequent, repeated furniture class.

Bottle

Tests small-object detection at different distances.

Keyboard

Tests recognition of specialized desk equipment.

Instead of passing only raw images to later components, the pipeline creates a structured observation for every detected object.

{
  "timestamp": "2026-09-07T10:15:20.000Z",
  "camera_id": "camera_1",
  "class": "laptop",
  "confidence": 0.94,
  "bounding_box": [120, 80, 420, 350],
  "mask_reference": "camera_1_20260907_101520_laptop_01"
}
Important limitation: A detection such as “laptop” identifies an object category, not a specific physical device such as “Laptop #3.” Persistent identity must be added later through temporal tracking, cross-camera matching, visual features, or an external identifier.
YOLO Object Detection

Recognizes multiple objects in one frame and returns a class label, confidence score, and bounding box for each detection.

Pretrained First

Allows early testing on common objects before collecting a large custom dataset.

Custom Fine-Tuning

Can be introduced if pretrained classes do not recognize specialized classroom equipment reliably.

SAM 3 Refinement

Uses YOLO boxes as prompts and produces masks that preserve object shape more precisely than rectangles.

Local Processing

Keeps classroom data on the Linux host and avoids continuous cloud upload.

Resolution Scaling

Compares 720p and 1080p inputs to balance small-object detail against CPU latency.

A standard image-classification model can predict that a classroom image contains a laptop, but it does not identify every separate laptop or show where each one is located. YOLO performs both classification and localization, allowing the system to detect multiple objects in one frame and return an individual class label, confidence score, and bounding box for each object. This makes YOLO more suitable for classroom inventory, people counting, tracking, and scene comparison.

  • Capture a timestamped classroom frame and return a class label, confidence score, and bounding box for every accepted YOLO detection
  • Evaluate all six target classes using a separate held-out test set that was not used for training
  • Reach the project target of at least 80% correct-class detections on the held-out images and also report per-class precision, recall, and mAP50
  • Compare detection performance under bright light, reduced light, distance, unusual viewing angles, and partial occlusion
  • Pass YOLO bounding boxes to SAM 3 as prompt boxes and generate a binary mask for each selected object
  • Measure mask quality using Intersection over Union on a manually labeled validation sample
  • Measure end-to-end YOLO + SAM latency at 720p and 1080p on the actual Linux host
  • Complete one capture-and-processing cycle within 10 seconds, or document the optimization needed to reach that target
Current constraint: The primary Linux host does not have a dedicated NVIDIA GPU. All latency values must therefore be measured on the actual project hardware rather than treated as guaranteed model performance.
ResolutionWhat Will Be MeasuredExpected TradeoffStatus
720p (1280×720)YOLO latency, SAM latency, total cycle time, class metrics, mask IoUFaster processing but reduced detail for small or distant objectsTO TEST
1080p (1920×1080)YOLO latency, SAM latency, total cycle time, class metrics, mask IoUMore object detail with increased CPU and memory demandTO TEST
4K (3840×2160)Optional comparison if lower resolutions cannot preserve required detailHighest input detail but likely to exceed the CPU time budgetOPTIONAL
  • Baseline Test: Determine which target classes are already supported reliably by the pretrained YOLO model.
  • Dataset: Collect representative classroom images across both camera viewpoints and split them into training, validation, and held-out test sets.
  • Fine-Tuning: Train a custom model only for classes that do not meet the required detection performance.
  • Prompt Refinement: Test box prompts and optional positive/negative points when adjacent objects cause SAM mask leakage.
  • Optimization: Evaluate lower input resolution, model-size selection, ONNX Runtime, and quantization if the local pipeline exceeds 10 seconds.
Ultralytics Object Detection Documentation Ultralytics · Official Documentation
  • Explains object classes, confidence scores, bounding boxes, inference, training, and evaluation.
  • Supports the choice of YOLO as the first computer-vision stage.
TensorFlow Object Detection Tutorial TensorFlow · Official Documentation
  • Demonstrates pretrained object detectors and the tradeoff between speed and accuracy.
  • Provides background for comparing object-detection approaches.
Google AI Edge Object Detection Google AI Edge · Official Documentation
  • Describes transfer learning and lightweight local deployment.
  • Supports possible custom training for specialized classroom classes.