tahamajs/Deep_UT/CA3_Object_Detection - Storage Bucket

CA3: Object Detection

This assignment explores advanced computer vision techniques for object detection and segmentation, implementing two state-of-the-art architectures: Fast-SCNN for real-time semantic segmentation and Oriented R-CNN for arbitrary-oriented object detection.

Overview

The assignment consists of two specialized projects:

1. Fast-SCNN Implementation: Real-time semantic segmentation for mobile devices
2. Oriented R-CNN Implementation: Detection of objects with arbitrary orientations

Assignment Snapshot

• Primary Notebooks:
  • Fast_SCNN/code/NNDL_CA3_1.ipynb
  • Oriented_RCNN/code/NNDL_CA3_2.ipynb
• Datasets: Cityscapes subset for semantic segmentation and oriented aerial/document imagery (DOTA-style) for rotation-aware detection.
• Key Benchmarks: Fast-SCNN ≈ 0.62 mIoU with <1.2M params; Oriented R-CNN increases recall on rotated targets compared to axis-aligned baselines.

Contents

• Fast_SCNN/: Real-time semantic segmentation implementation
• Oriented_RCNN/: Arbitrary-oriented object detection system

Each subfolder contains:

• code/: PyTorch implementations
• description/: Assignment specifications
• paper/: Research papers and references
• report/: Detailed analysis and results

Fast-SCNN: Real-Time Semantic Segmentation

Key Features

• Efficient Architecture: Depthwise separable convolutions for computational efficiency
• Multi-Scale Processing: Pyramid pooling module for global context
• Real-Time Performance: Optimized for mobile and embedded deployment
• Urban Scene Understanding: Segmentation of roads, buildings, vehicles, pedestrians

Technical Details

• Learning to Downsample: Initial downsampling module with skip connections
• Global Feature Extractor: Pyramid pooling with multiple kernel sizes (1×1, 2×2, 3×3, 6×6)
• Feature Fusion: Concatenation of multi-scale features with channel attention
• Loss Function: Cross-entropy with class balancing for imbalanced segmentation

Results

• IoU Score: 0.62 average across all classes
• Model Size: 1.2M parameters (vs. 50M+ for standard segmentation models)
• Inference Speed: 30+ FPS on mobile GPUs
• Memory Efficiency: 50MB model size suitable for edge deployment

Visual	Description
	Qualitative segmentation masks on Cityscapes samples.
	Multi-class predictions illustrating fine-grained boundaries.

Oriented R-CNN: Arbitrary-Oriented Object Detection

Key Features

• Oriented Anchors: 5-parameter anchor representation (x, y, w, h, θ)
• Rotated ROI Align: Rotation-aware feature extraction
• Geometric Transformations: Proper handling of oriented bounding boxes
• IoU Computation: Specialized intersection-over-union for rotated rectangles

Technical Details

• Region Proposal Network (RPN): Oriented anchor generation and classification
• Rotated RoI Align: Bilinear sampling with rotation compensation
• Bounding Box Regression: 5-parameter regression (dx, dy, dw, dh, dθ)
• Orientation Encoding: Angle representation and normalization

Results

• Detection Accuracy: Superior performance on oriented objects vs. axis-aligned methods
• Geometric Precision: Accurate localization of rotated objects
• Robustness: Handles various orientations and aspect ratios
• Application: Ship detection in satellite imagery, text detection in documents

Visual	Description
	Rotation-aware bounding boxes on aerial datasets.
	Tight localization highlighting angle regression quality.

Key Concepts Demonstrated

Semantic Segmentation

• Pixel-wise Classification: Assigning class labels to every pixel
• Encoder-Decoder Architectures: U-Net style networks for segmentation
• Efficient Convolutions: Depthwise separable and grouped convolutions
• Multi-Scale Features: Pyramid pooling and feature fusion

Object Detection

• Region Proposal Networks: Generating candidate object locations
• RoI Pooling/Align: Feature extraction from proposed regions
• Bounding Box Regression: Refining object localization
• Orientation Handling: 5-DoF bounding box representation

Geometric Transformations

• Rotation Matrices: Mathematical representation of orientations
• IoU for Rotated Boxes: Intersection computation for oriented rectangles
• Spatial Transformations: Affine and projective transformations
• Coordinate Systems: Converting between different geometric representations

Educational Value

This assignment provides deep understanding of:

• Advanced CV Architectures: State-of-the-art detection and segmentation models
• Geometric Computer Vision: Handling rotations and orientations
• Efficient Deep Learning: Model compression and mobile optimization
• Real-world Applications: Aerial imagery, autonomous driving, document analysis

Dependencies

• Python 3.8+
• PyTorch 1.9+
• torchvision
• OpenCV
• NumPy, Pandas
• Matplotlib, Seaborn
• CUDA (recommended for training)

Usage

Fast-SCNN

1. Navigate to Fast_SCNN/code/
2. Run data preprocessing and model training
3. Evaluate segmentation performance on test sets
4. Review results in Fast_SCNN/report/

Oriented R-CNN

1. Navigate to Oriented_RCNN/code/
2. Execute the training pipeline for oriented detection
3. Test on oriented object datasets
4. Analyze geometric accuracy in Oriented_RCNN/report/

References

• Fast-SCNN Description
• Oriented R-CNN Description
• Research Papers | Research Papers
• Implementation Reports | Implementation Reports

---

Course: Neural Networks and Deep Learning (CA3) Institution: University of Tehran Date: September 2025 /Users/tahamajs/Documents/uni/LLM/Deep_UT/CA1_Neural_Networks_Basics/README.md