Object Detection Models Comparison

---

👥 Team Members

1. Mohamed Hassan
2. Omar Hany
3. Mohamed Mohamed Ibrahim

---

1. Model Evaluation: Faster R-CNN ResNet-101 (1024×1024)

Model Architecture

Faster R-CNN (ResNet-101 backbone) with 1024×1024 input. A two-stage detector consisting of:

• Backbone (ResNet-101): Deep backbone for high-level semantic feature extraction.
• RPN: Generates candidate object regions via anchors.
• RoI Align: Resizes regions to fixed resolution, preserving spatial alignment.
• Detection Head: Classifies regions and refines bounding boxes.

Quantitative Results

COCO 2017 Validation

Metric	Value	Metric	Value
mAP @[0.50:0.95]	0.317	AP @ 0.50	0.456
Avg IoU (Matched)	0.830	AP (Small/Med/Lrg)	0.127 / 0.347 / 0.498

Pascal VOC 2007 (COCO metrics)

Metric	Value	Metric	Value
mAP @[0.50:0.95]	0.524	AP @ 0.50	0.744
Avg IoU (Matched)	0.841	AP (Small/Med/Lrg)	0.168 / 0.407 / 0.619

Analysis

• Performance: mAP 0.317 on COCO; significantly better on VOC (0.524).
• Localization: High Avg IoU (~0.83) confirms the two-stage refinement is highly effective.
• Scale: Performs well on Medium/Large objects but struggles with Small objects.

Qualitative Analysis

Successful Detections:

Failure Cases:

Feature Maps:

---

2. Model Evaluation: SSDLite320 MobileNetV3-Large

Model Architecture

SSDLite320 (MobileNetV3-Large backbone). A single-stage, efficient detector for edge devices.

• Backbone: Lightweight MobileNetV3 with SE modules.
• SSD: Direct bounding box prediction without region proposals.
• Resolution: Fixed 320×320 input for speed, limiting small object visibility.

Quantitative Results

COCO 2017 Validation

Metric	Value
mAP @[0.50:0.95]	0.2107
AP @ 0.50	0.3388
AP @ 0.75	0.2190
AP (Small)	0.0034
AP (Medium)	0.0947
AP (Large)	0.4119
Mean IoU (>0.7)	0.8062

Pascal VOC 2007 (COCO metrics)

Metric	Value
mAP @[0.50:0.95]	0.4150
AP @ 0.50	0.6364
AP @ 0.75	0.4443
AP (Small)	0.0056
AP (Medium)	0.1784
AP (Large)	0.5668

Analysis

• Trade-off: The mAP (0.4150) reflects the efficiency trade-off.
• Small Object Blindness: AP (Small) is extremely low (0.0056) due to 320px downsampling.
• Strengths: Competent on large, prominent objects (AP Large 0.5668).

Qualitative Analysis

Successful Detections:

Failure Cases:

Feature Maps:

---

3. Model Evaluation: YOLOv12-X

Model Architecture

YOLOv12: High-speed detector using attention mechanisms (FlashAttention, Area Attention) and a hierarchical backbone (R-ELAN).

Quantitative Results

Efficiency: 59.1M Params, 199.0 GFLOPs, 11.79 ms/img latency.

Detection Performance

Dataset	mAP @[0.50:0.95]	Avg IoU	mAP @ 0.50
COCO 2017	0.718	0.554	0.841
Pascal VOC 2007	0.898	0.714	0.888

Qualitative Analysis

Successful Cases (COCO & Pascal):

Failure Cases (COCO & Pascal):

Feature Map Visualization: