jeff-calderon/Aerial_Car_Roboflow

Car_Ariel640 Dataset

File Structure

.
├── README.md
├── split_dataset.py
├── analyze_vehicle_stats.py
├── analyze_real_colors.py
├── format_stats.py
├── format_real_colors.py
├── vehicle_stats.json
├── real_vehicle_colors.json
├── train/
│   ├── _annotations.coco.json
│   ├── *.jpg (4,752 files)
├── valid/
│   ├── _annotations.coco.json
│   ├── *.jpg (1,578 files)
├── test/
│   ├── _annotations.coco.json
│   ├── *.jpg (694 files)

Analysis Scripts:

• split_dataset.py: Dataset splitting utility (train/valid/test)
• analyze_vehicle_stats.py: Computes orientation and basic mask-based stats
• analyze_real_colors.py: (New) Analyzes real-world vehicle colors using expanded bounding boxes
• format_real_colors.py: Formats real-world color analysis for documentation
• vehicle_stats.json / real_vehicle_colors.json: JSON output of analysis runs

Dataset Specifications

• Image Dimensions: 640 × 640 pixels (verified across all splits)
• Format: COCO JSON annotations
• Task: Instance Segmentation
• Categories: Large Vehicle (ID 6), Small Vehicle (ID 10)

Dataset Audit: Instance Segmentation for Vehicles

Objective: Sanity check on COCO JSON annotation files to ensure correct data distribution for "large_vehicle" and "small_vehicle".

1. Category Identification

• Large Vehicle: ID 6
• Small Vehicle: ID 10

2. Instance Counts Summary

Split	Large Vehicles	Small Vehicles	Total Instances
Train	16,798	116,074	132,872
Valid	4,955	33,298	38,253
Test	2,728	14,057	16,785
TOTAL	24,481	163,429	187,910

Distribution: The dataset is skewed towards small vehicles (~87% small vs. ~13% large).

3. Size Logic & Area Validation

• Large Vehicles:
  • Mean Area: ~2,446 pixels² (Train)
  • Max Area: 230,400 pixels²
  • Observation: A significant portion (approx. 50%) of "large vehicles" fall into the COCO "small" category (< 1024 px²). This suggests the dataset's definition of "large" is semantic (class-based) rather than strictly based on COCO's absolute pixel area thresholds.
• Small Vehicles:
  • Mean Area: ~293 pixels² (Train)
  • Max Area: ~7,088 pixels²
  • Observation: The vast majority (>90%) are indeed "small" by COCO standards (< 1024 px²).

4. Segmentation & Data Integrity

• Polygon Format: All instances use standard polygon segmentation.
• Crowds: No instances are marked with iscrowd: 1.
• File Consistency: Verified. Every file listed in the JSON images array exists on disk.

5. Vehicle Color & Orientation Statistics

Analysis Method: 500 randomly sampled vehicle instances per dataset analyzed for pixel-level color and geometric properties.

Real-World Color Analysis (Expanded Bounding Boxes)

Analysis Note: Colors sampled from expanded bounding box regions (1.3x) to capture real-world vehicle appearance, including body, roof, and immediate context, with atmospheric artifact filtering.

RGB Statistics (8-bit):

Dataset	Vehicle Type	Mean R,G,B	Median R,G,B
Train	Small Vehicle	98.4, 97.5, 98.2	94.8, 94.0, 94.9
	Large Vehicle	121.0, 118.5, 117.0	125.3, 122.8, 121.3
Test	Small Vehicle	96.2, 95.4, 96.3	92.3, 91.4, 92.3
	Large Vehicle	122.4, 121.0, 119.7	129.3, 128.0, 126.2
Valid	Small Vehicle	96.0, 95.5, 97.7	92.4, 91.8, 94.0
	Large Vehicle	132.7, 131.0, 126.7	138.9, 136.9, 130.8

HSV Statistics:

Dataset	Vehicle Type	Mean H,S,V	Dominant Colors
Train	Small Vehicle	163.9°, 12.6%, 40.5%	Cyan-Blue (72.9%), Green (14.5%)
	Large Vehicle	138.5°, 15.2%, 49.0%	Cyan-Blue (62.7%), Magenta (19.8%)
Test	Small Vehicle	163.3°, 13.6%, 39.8%	Cyan-Blue (68.5%), Green (16.1%)
	Large Vehicle	139.6°, 13.2%, 49.7%	Cyan-Blue (64.4%), Green (20.1%)
Valid	Small Vehicle	175.2°, 13.9%, 40.0%	Cyan-Blue (72.8%), Magenta (14.5%)
	Large Vehicle	145.1°, 14.7%, 53.7%	Cyan-Blue (71.7%), Green (15.7%)

Rotated Minimum Area Rectangle Analysis

Geometric Properties:

Dataset	Vehicle Type	Mean Angle	Aspect Ratio	Orientation Pattern
Train	Small Vehicle	3.1° ± 55.4°	2.17 ± 0.56	Mixed (45.2° mean orientation)
	Large Vehicle	-2.4° ± 56.3°	3.27 ± 1.15	Mixed (47.4° mean orientation)
Test	Small Vehicle	-0.2° ± 55.3°	2.16 ± 0.59	Mixed (44.4° mean orientation)
	Large Vehicle	8.3° ± 53.4°	3.46 ± 1.53	Mixed (43.6° mean orientation)
Valid	Small Vehicle	1.6° ± 56.7°	2.21 ± 0.82	Mixed (47.3° mean orientation)
	Large Vehicle	0.3° ± 53.6°	3.52 ± 2.26	Mixed (44.2° mean orientation)

Key Insights (Real-World)

Real-World Color Patterns:

• Cyan-Blue Dominance: ~70% of vehicles appear in the blue-gray spectrum, reflecting realistic aerial imagery conditions (atmospheric scattering)
• Brightness Distinction: Large vehicles are consistently brighter (RGB ~120-130) than small vehicles (RGB ~96-98), likely due to larger reflective surface areas
• Low Saturation: Very low saturation (12-15%) across all vehicle types confirms the muted, realistic appearance of objects in aerial views
• Consistency: Color distributions are highly consistent across train, test, and validation sets, ensuring reliable model training

Orientation Patterns:

• Random distribution: Nearly uniform angle distribution across all datasets, indicating vehicles are randomly oriented in aerial imagery
• Aspect ratio differentiation: Large vehicles consistently show higher aspect ratios (~3.3 vs ~2.2), confirming elongated shape characteristics
• No systematic bias: Mean angles near 0° with high standard deviation (~55°) suggests no preferred parking/driving orientation

Updated Methodology (Real-World Colors)

Color Analysis (Real Vehicle Colors):

1. Bounding Box Expansion: Original bbox expanded 1.3x to capture complete vehicle including body, roof, and immediate context
2. Smart Pixel Sampling: 70% core vehicle pixels + 30% border region for comprehensive color representation
3. Atmospheric Filtering: Excludes obvious artifacts (sky V>200, shadows V<30, oversaturated S<10) using HSV thresholds
4. Color Space Conversion: RGB → HSV for perceptual uniformity and hue-based color naming
5. Statistical Aggregation: Mean/median computed across 500 sampled instances per dataset with seed=42 for reproducibility

Rotated Rectangle Analysis:

1. Contour Extraction: Polygon vertices converted to OpenCV contour format
2. Minimum Area Rectangle: cv2.minAreaRect() computes optimal rotated bounding box
3. Angle Normalization: Angles normalized to [-90°, 90°] range with width as longer side
4. Aspect Ratio: Width/Height ratio indicating elongation (higher = more rectangular)
5. Orientation: Absolute angle indicating deviation from horizontal axis

Sampling Strategy:

• Random selection of 500 vehicle instances per dataset (seed=42 for reproducibility)
• Balanced representation across small vehicles (ID: 10) and large vehicles (ID: 6)
• Real-world colors captured from original aerial imagery, not segmentation masks