Vision-Dissect

Systematic benchmarking and layer-by-layer analysis of computer vision models for depth estimation, object detection, segmentation, and pose estimation.

Benchmark Results (November 11, 2025)

Model	Inference (s)	FPS	Memory (MB)	Parameters	Layers
DepthPro	14.811 ± 1.201	0.07	3643	952M	57
Depth Anything V2 Small	0.058 ± 0.047	17.13	105	24.8M	33
Depth Anything V2 Base	0.111 ± 0.055	9.02	381	97.5M	33
Depth Anything V2 Large	0.262 ± 0.009	3.82	1290	335M	33
YOLO11n-Detect	0.178 ± 0.283	5.63	19	2.62M	89
YOLO11n-Segment	0.071 ± 0.062	14.16	21	2.87M	102
YOLO11n-Pose	0.089 ± 0.102	11.21	20	2.87M	98
MobileSAM	7.238 ± 2.302	0.14	67	10.1M	138

Hardware: NVIDIA GeForce RTX 3060 Laptop GPU, CUDA 12.6, PyTorch 2.9.0+cu126

Key Findings

• Depth Anything V2 Small is 255× faster than DepthPro (0.058s vs 14.8s)
• YOLO11n-Segment outperforms Detection variant by 2.5× (14.16 FPS vs 5.63 FPS)
• Zero sparsity across all 289 dissected layers indicates excellent training efficiency
• Depth Anything V2 Base has sign encoding bug (positive depth values require flip)
• Detection variant suffers 159% coefficient of variation (GPU scheduling issues)

Models Analyzed

Depth Estimation

• DepthPro (Apple) - 57 Conv/ConvTranspose layers
• Depth Anything V2 Small/Base/Large - 33 Conv2d layers each

Object Detection & Segmentation

• YOLO11n-Detect - 89 Conv2d layers
• YOLO11n-Segment - 102 Conv2d/ConvTranspose2d layers
• YOLO11n-Pose - 98 Conv2d layers
• MobileSAM - 138 layers

Structure

vision-bench/          # Benchmark framework
  unified_benchmark.py # Main benchmark script
  results/             # Performance metrics & reports
  viz/                 # Layer visualizations (PNG + NPY)

docs/                  # Detailed analysis reports
  MODEL_ANALYSIS.md    # Cross-model comparison
  DEPTHPRO_DETAILED.md # DepthPro findings
  DEPTH_ANYTHING_V2_DETAILED.md # Depth Anything analysis
  YOLO11_FAMILY_DETAILED.md     # YOLO11 variants

apps/                  # Interactive applications
server/                # FastAPI model serving
compare/               # Comparison utilities
dissect/               # Model dissection tools
explore/               # Experimentation scripts

Quick Start

# Install dependencies
pip install -e .

# Run comprehensive benchmark
python vision-bench/unified_benchmark.py

# Dissect specific model
python dissect/dissect_depthpro.py
python dissect/dissect_depthanything.py

# Launch interactive app
streamlit run apps/depthpro_app.py

Web Explorer

A professional Next.js web interface for exploring benchmark results:

cd explorer
npm install
npm run dev

Open http://localhost:3000

Features:

• Benchmarks: Performance metrics comparison
• Layer Visualizations: Detailed layer analysis with statistics
• Graph View: Interactive computational graph
• Live Monitor: Real-time benchmark progress

Deployment: Ready for Vercel with automatic GitHub data source

• See explorer/README.md for setup
• See explorer/DATA_SOURCE.md for technical details

Visualizations

All layer dissection visualizations saved as:

• PNG: 4×4 grid of 16 channels (visual inspection)
• NPY: 8 channels (numerical analysis)

Documentation

See docs/ for comprehensive analysis:

• MODEL_ANALYSIS.md - Cross-model benchmark comparison
• DEPTHPRO_DETAILED.md - 57-layer dissection, extreme activation analysis
• DEPTH_ANYTHING_V2_DETAILED.md - 3-variant comparison, sign bug documentation
• YOLO11_FAMILY_DETAILED.md - 89/102/98 layer analysis per variant

Requirements

• Python 3.11+
• PyTorch 2.9+ with CUDA 12.6
• transformers (DepthPro)
• ultralytics (YOLO11)
• onnx / onnxruntime
• opencv-python
• numpy, matplotlib

Citation

Repository: Vision-Dissect
Owner: infiniV
Branch: main
Benchmark Date: November 11, 2025