Canadian University Dubai
School of Engineering, Applied Science and Technology
Instructor: Dr. Najla Al Futaisi
BCS 407 Artificial Intelligence Spring 2026

Campus Infrastructure
Object Detection
with YOLOv11n

Group Members
Tahamid Hossain 20220001801
Sumaid Bin Omar 20220001454
Parth Aggarwal 20220001200
Rufaid Bin Omar 20230003171
Arham Bin Azad 20220001121
 YOLOv11n · 2.6M params CUDA · RTX 4060 Batch 04 · 2026-04-24 100 Epochs

Campus Infrastructure
Object Detection
with YOLOv11n

End-to-end pipeline for detecting four campus assets — projectors, whiteboards, fire extinguishers, and door signs — using a curated 800-image dataset with perfect class balance, achieving near-saturating accuracy on safety-critical classes.

Comparative Analysis Batch 04 vs Batch 05 · updated custom dataset YOLOv11s Comparative Analysis Batch 04 vs 05 vs 06 · baseline + dataset + model uplift
0%
mAP@0.5
Test set · 80 images
0%
mAP@0.5:0.95
Tight IoU eval
1.000
Precision
Zero false positives
0%
Recall
Macro average
01 · Pipeline

End-to-End Workflow

Seven notebooks take raw dataset exports through aggregation, splitting, health-checking, training, evaluation, inference, and export — each stage's artefacts feeding the next.

NB01
Aggregate
200 (image, label) pairs per class from Roboflow/Kaggle exports → JPEG uniformity
→
NB02
Remap + Split
Global ID assignment · class ID normalisation · stratified 70/20/10 split
→
NB03
Health Check
Class balance, box distribution, leakage detection, dimension stats
→
NB04
Train YOLOv11n
100 epochs · SGD · AMP FP16 · mosaic augmentation · RTX 4060
→
NB05
Evaluate
Held-out test split · mAP, P/R curves · confusion matrix · qualitative preds
→
NB06
Live Inference
Single image · batch · video · webcam modes against best.pt
→
NB07
ONNX Export
Opset 12 · dynamic axes · onnx-simplifier · 50% size reduction
↗
02 · Dataset

Data Collection & Preparation

2.1 · Data Collection — NB01 ▶

Raw images sourced from Roboflow, except doorsign2–4, which were custom-captured from HUB campus door signs, annotated, and prepared by the team. NB01 aggregated exactly 200 (image, label) pairs per class into data/aggregated/<class>/, re-encoding all images to JPEG.

Class Available Pairs After Cap Source
● projector 319 200 Projector1, 2, 3 (Roboflow)
● whiteboard 200 200 Single Roboflow export
● fire_extinguisher 848 200 Kaggle train/valid/test
● door_sign 240 200 doorsign1 (Roboflow); doorsign2–4 (custom — HUB campus, annotated by team)
2.2 · Stratified Split — NB02 ▶

NB02 merged all four classes into a unified Ultralytics-compatible dataset with global ID re-indexing and class ID normalisation (0=projector, 1=whiteboard, 2=fire_extinguisher, 3=door_sign). Stratified 140/40/20 split per class.

Class Train Val Test Boxes/img (train)
● projector 140 40 20 0.82
● whiteboard 140 40 20 1.09
● fire_extinguisher 140 40 20 1.19
● door_sign 140 40 20 1.86
Total 560 160 80 —
💡
door_sign averages ~1.9 boxes/image across all splits — exit + number placards commonly co-occur. projector falls below 1.0 boxes/image, indicating more background frames.
Spot-check grid of class samples
Figure 1. 3×4 spot-check grid of class samples drawn from the split dataset.
2.3 · Dataset Health Check — NB03 ▶
Train images560
Val images160
Test images80
Total boxes997
Empty labels4.4%
Tiny boxes0
Leakage0 dups
Median dim640×640
Class distribution bar chart
Figure 2. Per-class image counts by split.
Box area histogram
Figure 3. Normalised box area distribution per class.
Image dimension scatter
Figure 4. Width vs height scatter — 800 images. Cluster at (640,640) dominates; sparse high-res outliers from Kaggle fire extinguisher photos.
03 · Methodology

Model Selection & Training Config

Architecture Comparison

Architecture Params COCO mAP Latency (T4) Verdict
Faster R-CNN (ResNet-50) ~41 M 42.9 ~47 ms Too slow
SSD MobileNetV2 ~3.4 M 22.1 ~1.2 ms Low ceiling
YOLOv8n 3.2 M 37.3 ~1.47 ms Strong baseline
YOLOv11n 2.6 M 39.5 ~1.55 ms ✓ Selected
YOLOv11s 9.4 M 47.0 ~2.46 ms +acc, +params
YOLOv11m 20.1 M 51.5 ~4.70 ms Not edge
🎯
Why YOLOv11n? 2.6M params — 0.6M fewer than YOLOv8n with 2.2 points higher COCO accuracy. C3k2 + SPPF backbone with depthwise convolutions exports cleanly to ONNX. Transfer learning from COCO provides critical feature initialisation for our small 560-image training set.
Training Configuration ▶
Weightsyolo11n.pt
Epochs100
Patience15
Batch size16
Image size640×640
OptimizerSGD
lr0 / lrf0.01 / 0.01
Warmup3 epochs
AMPTrue (FP16)
Seed42
IoU thresh0.7
DeviceCUDA:0

Loss Weights

Box (CIoU)7.5
Classification (BCE)0.5
DFL1.5
Data Augmentation ▶
Mosaic1.0 (off last 10 ep)
RandAugmentEnabled
Random Erasing0.4 prob
Horizontal flip0.5 prob
Scale jitter±50%
Translation±10%
HSV hue0.015
HSV sat/val0.7 / 0.4
MixUp/CutMixDisabled
Vertical flipDisabled
🔀
Mosaic concatenates 4 training images into a single 640×640 tile — effectively quadrupling context diversity and forcing the model to handle partially visible objects. Disabled for the final 10 epochs to consolidate clean features.
04 · Architecture

YOLOv11n — Model Structure

Single-stage anchor-free detector with task-aligned assignment (TAL) — eliminates anchor tuning and improves small-object recall.

Input
RGB Image · 640×640
↓
── Backbone ──
C3k2 Blocks
Compact CSP bottleneck
2×depthwise conv
SPPF
Spatial Pyramid Pooling – Fast
Multi-scale context
↓
── Neck ──
PANet
Bidirectional feature fusion
P3/8 · P4/16 · P5/32
↓
── Head ──
Classification Branch
BCE loss · 4 classes
Regression Branch
DFL · CIoU · anchor-free
↓
TAL Assignments → NMS → Detections
Task-Aligned Learning · IoU threshold 0.7
Parameters2.6 M
ParadigmOne-stage · anchor-free
COCO mAP@0.5:0.9539.5
T4 latency~1.55 ms/img
05 · Training

Training Curves & Convergence

Model peaked at epoch 60 (val mAP@0.5 = 0.9489). Val box loss flattened around 0.84 while train loss continued decreasing — mild overfitting, correctly halted by patience=15.

Val mAP over Epochs
Box Loss over Epochs
Epoch Train Box Loss Val mAP@0.5 Val mAP@0.5:0.95
1 1.068 0.199 0.122
10 ~0.88 ~0.650 ~0.450
30 ~0.77 ~0.880 ~0.680
60 (best) 0.744 0.9489 0.7251
100 (final) 0.511 0.938 0.735
Training curves
Figure 6. Training curves — box + classification loss on train/val (left); mAP@0.5 and mAP@0.5:0.95 on val set across 100 epochs (right).
Label distribution
Figure 5. YOLO-generated label distribution: class frequency (left) and bounding-box spatial heatmap (right).

Validation Batch — Ground Truth vs. Predictions

Ground Truth Val batch 0 labels
Predictions Val batch 0 predictions
Ground Truth Val batch 1 labels
Predictions Val batch 1 predictions
Ground Truth Val batch 2 labels
Predictions Val batch 2 predictions

Training Batch Mosaic

Training batch mosaic
Figure 7. Example mosaic training batch (batch 0) — augmented 4-tile compositions at 640×640.

Post-Training Sanity Check

Sanity predictions grid
Figure 9. 1×4 sanity inference grid — correct class assignment with high confidence scores confirmed.
06 · Results

Test-Set Evaluation

Evaluated exclusively on the held-out test split (80 images, 102 boxes) using best.pt at conf=0.25, IoU=0.5.

0%
mAP@0.5
0%
mAP@0.5:0.95
100%
Precision (macro)
0%
Recall (macro)

Per-Class Performance

● projector
P: 1.000 R: 0.804 mAP@0.5: 0.896
mAP@0.5
89.6%
mAP@0.5:0.95
72.9%
Recall
80.4%
● whiteboard
P: 1.000 R: 0.763 mAP@0.5: 0.876
mAP@0.5
87.6%
mAP@0.5:0.95
77.7%
Recall
76.3%
● fire_extinguisher
P: 1.000 R: 0.957 mAP@0.5: 0.978
mAP@0.5
97.8%
mAP@0.5:0.95
90.5%
Recall
95.7%
● door_sign
P: 1.000 R: 0.922 mAP@0.5: 0.982
mAP@0.5
98.2%
mAP@0.5:0.95
77.3%
Recall
92.2%

Confusion Matrices

Confusion matrix raw counts
Figure 11. Raw-count confusion matrix (4×4 + background).
Confusion matrix normalised
Figure 12. Row-normalised confusion matrix — diagonal dominance confirms class separability.
Custom confusion matrix
Figure 13. Custom dual-panel Seaborn confusion matrix.

Precision–Recall & Threshold Curves

PR curve
Figure 14. Precision–Recall curves per class.
F1 curve
Figure 17. F1 score vs. confidence threshold.
Precision vs confidence
Figure 15. Precision vs. confidence — all classes at P=1.0 above conf≈0.4.
Recall vs confidence
Figure 16. Recall vs. confidence threshold.

Qualitative Predictions

Qualitative predictions grid
Figure 18. 2×4 grid of test-set images with predicted bounding boxes and confidence scores overlaid.

Ultralytics Results Dashboard

Ultralytics results dashboard
Figure 10. 10-subplot training dashboard — loss curves, precision, recall, mAP across 100 epochs.
07 · Discussion

Strengths & Limitations

✓ Zero False Positives
Macro precision = 1.0 across all classes. The model never hallucinates detections above the 0.25 threshold — critical for inventory and safety-audit applications where false alarms erode trust.
✓ Strong mAP@0.5
93.3% mAP@0.5 on 80 test images with a 2.6M parameter model. Fine-tuning on a narrow four-class domain closes the gap from COCO's 39.5% dramatically.
✓ Safety-Critical Near-Saturation
fire_extinguisher (97.8%) and door_sign (98.2%) mAP@0.5 — the highest-value outcome for compliance auditing. Visually distinctive appearance aids detection.
✓ Edge-Ready Weight Size
22 MB PyTorch → 11 MB ONNX. ONNX export enables TensorRT, OpenVINO, and CoreML deployment without a PyTorch runtime dependency.
⚠ Projector/Whiteboard Recall
Recall of 0.804 / 0.763 respectively. High intra-class variance: ceiling-mounted vs. desktop projectors, reflective whiteboards with/without text. Lowest boxes/image ratio (~0.82).
⚠ mAP@0.5:0.95 Gap
13.7-point gap between mAP@0.5 (0.933) and mAP@0.5:0.95 (0.796). The lightweight regression head lacks capacity for sub-pixel box refinement. YOLOv11s or cascaded refinement would close this.
⚠ Small Dataset
560 training images for 4 classes. 4.4% empty-label rate reduces effective positives. Scaling to ~500 images per class would likely push projector and whiteboard recall above 0.90.
⚠ Single Domain
All images from similar institutional settings. Performance on unusual campus layouts, diverse lighting, or non-standard equipment may degrade due to domain shift.

Improvement Directions

Issue Suggested Fix
Low projector/whiteboard recall 100–200 additional images per class from varied viewpoints and lighting
mAP@0.5:0.95 gap Upgrade to YOLOv11s; or add test-time augmentation (TTA)
Potential domain shift Add online images from different institutions; apply colour jitter at inference
Overfitting after epoch 60 Add dropout (currently 0.0) or stronger weight decay
08 · Deployment

Inference & Export

Live Inference Modes — NB06

🖼
Single Image
Load, infer, display detection with latency measurement
📁
Batch / Folder
Per-class counts and throughput (images/s) over directory
🎬
Video File
Frame-by-frame annotation + annotated video + preview grid
📷
Webcam
Live feed with real-time overlay; stop via widget button
⚙
CONF_THRESH = 0.25 · IMGSZ = 640 · device auto-selected (CUDA → MPS → CPU)

ONNX Export — NB07

ONNX opset12
Dynamic axesTrue
SimplifyTrue
Size reduction~50%
.pt
best.pt
PyTorch checkpoint · full training state
~22 MB
ONNX
best.onnx
Framework-agnostic · TensorRT / OpenVINO / CoreML
~11 MB
Conclusion

A complete object detection pipeline for four campus infrastructure classes was built and evaluated. YOLOv11n achieves mAP@0.5 = 93.3% with zero false positives at the chosen confidence threshold. Safety-critical classes fire_extinguisher and door_sign are detected at near-saturation accuracy (mAP@0.5 > 97.8%). Both PyTorch and ONNX weights are production-ready for real-time campus deployment.