EVOLVE

Extreme Vision Over Low-light and Volatile Environments

Master 2 SISE Project - Computer Vision
Université Lumière Lyon 2 | 2025–2026

---

Overview

EVOLVE is a structured object detection study focused on visually challenging real-world environments, using live metal concert imagery as a representative challenging domain.

Live metal concert imagery is used as a representative domain combining multiple simultaneous visual challenges:

• Low and highly uneven illumination
• Rapid chromatic shifts (red / blue dominant lighting)
• Motion blur
• High crowd density
• Severe occlusion
• Camera instability

These conditions generate strong variability in scene-level visual properties.

Beyond measuring detection performance under these conditions, EVOLVE investigates the following question:

To what extent can variations in object detection performance be explained by measurable scene-level properties?

The project aims to quantify the relationship between environmental descriptors (e.g., luminance, blur, density) and detection metrics.

---

Research Orientation

EVOLVE EVOLVE integrates:

• Structured dataset construction
• Annotation calibration and agreement analysis
• Controlled luminance stratification
• Quantitative performance modeling

The emphasis is placed on:

• Experimental control
• Statistical rigor
• Measurement transparency
• Reproducibility

Annotation is treated as a measurable experimental component rather than a preprocessing formality.

---

Dataset Construction

The dataset is built through a multi-stage controlled pipeline:

Raw video collection
↓
Frame extraction
↓
Luminance computation
↓
Stratified sampling
↓
Calibration subset creation
↓
Manual annotation (CVAT)
↓
Quality control
↓
Train / validation split

Full details are documented in:

• doc/dataset_pipeline.md
• doc/calibration_protocol.md
• doc/inter_annotator_analysis.md

Data Sources

• Personal concert footage
• Public YouTube videos (research use only)

The dataset is not distributed due to licensing constraints.

Dataset Statistics

Split	Images
Train	560
Validation	70
Test	70

Total images: 700

Class Distribution (Training Set)

Class	Instances
amp	744
guitar	440
drums	287
micro	275
mosh_pit	83
hands_raised	498

Global Class Distribution (All Splits)

Class	Instances
amp	879
guitar	494
drums	339
micro	329
mosh_pit	99
hands_raised	593

---

Target Classes

Class	Description
amp	Stage amplifiers or monitors
guitar	Guitar or bass instruments
drums	Drum kit or drum elements
micro	Vocal microphones
mosh_pit	Collective dynamic crowd movement
hands_raised	Clusters of raised arms

The mosh_pit class represents a collective behavioral phenomenon rather than a discrete object, introducing structural annotation complexity.

---

Modeling

• Framework: PyTorch
• Detector: YOLOv8 (Ultralytics)
• Training approach: transfer learning from COCO pre-trained weights
• Controlled comparison: pretrained vs training-from-scratch
• Evaluation stratified by luminance categories

---

Evaluation Strategy

Quantitative Metrics

• mAP (50 and 50-95)
• Precision / Recall per class
• IoU
• Stratified performance comparison

Qualitative Analysis

• Failure case inspection
• Lighting-dependent degradation
• Dense-scene ambiguity
• Class-specific instability patterns

---

Key Experimental Results

Validation Performance

Model	mAP@0.5:0.95	mAP@0.5
Pretrained (COCO)	~0.17	~0.30
Scratch	~0.03	~0.08

Transfer learning yields a substantial performance gain under extreme lighting conditions.

Training from scratch fails to converge to competitive detection performance, highlighting the importance of prior visual knowledge in low-light, high-variance domains.

Scene-Level Correlation Analysis (Pretrained Model)

Variable	r	R²	p-value	Interpretation
Mean luminance	0.391	0.153	0.001	Moderate positive correlation
Blur	-0.110	0.012	0.364	Not significant
Density	-0.080	0.006	0.511	Not significant
Occupied area	-0.070	0.005	0.567	Not significant

Detection recall is significantly associated with scene luminance (p = 0.001), whereas blur and structural density do not exhibit statistically significant influence in this dataset.

Qualitative Comparison

Example detection output on a high-illumination but high-glare frame:

The pretrained model correctly detects drums and guitar elements under extreme brightness and glare conditions, whereas the scratch model exhibits reduced localization accuracy and missed detections.

Main Takeaway

Object detection performance in extreme concert environments is not uniformly degraded. Instead, performance variability is partially structured and statistically linked to scene luminance.

This suggests that environmental descriptors can help anticipate detection reliability in real-world low-light deployments.

---

Repository Structure

EVOLVE/
├── configs/
├── data/
│ ├── raw/
│ ├── interim/
│ └── processed/
├── doc/
├── metadata/
├── notebooks/
├── results/
├── scripts/
└── README.md

---

Limitations

• Limited dataset size
• Domain specificity
• Residual annotation subjectivity
• Extreme lighting variability

These constraints are explicitly considered in the analysis.

---

Perspective

Future work may include:

• Dataset extension
• Domain adaptation experiments
• Lighting-aware preprocessing ablation
• Cross-architecture comparison

---

License

Academic project - non-commercial use only.

---

EVOLVE - Extreme Vision Under Extreme Conditions