How yaubernice/image-similarity works
Image Similarity is a visual-search studio and CLI for comparing deep image descriptors with per-model nearest-neighbour indexes. The main workflow is to manage a photo library, build an index for a selected model, submit an image or CLIP text query, and inspect ranked results; optional processing adds whitening, query expansion, and geometric reranking. Separate branches support Caltech-101 evaluation and DINOv2 fine-tuning. Some connections are documented capabilities rather than fully visible call paths in the excerpts, so those edges cite the README.
Main entry points: evaluate.py Retrieval evaluation · app.py Studio API · extractors.py Feature extractors · image_similarity.py CLI build and query
Components 15
Search studio
- Studio interface
static/app.js - Studio API
app.py
Retrieval runtime
- Feature extractors
extractors.py - Retrieval processing
retrieval.py - CLI build and query
image_similarity.py
Evaluation and training
- Retrieval evaluation
evaluate.py - Ranking metrics
metrics.py - DINOv2 fine-tuning
finetune.py
Image data
- Photo library
images/database - Model indexes
index - Caltech subset preparation
scripts/prepare_caltech101.py - Caltech queries and labels
- Fine-tuned checkpoint
- Evaluation outputs
results
Other components
- Studio user
Connections 18
- Studio user → Studio interface: uses
README.md - Studio interface → Studio API: calls API
static/app.js - Studio API → Photo library: manages photos
README.md - Studio API → Feature extractors: embeds queries
app.py - Studio API → Retrieval processing: applies whitening
app.py - Studio API → Model indexes: builds and searches
README.md - CLI build and query → Feature extractors: embeds images
image_similarity.py - CLI build and query → Retrieval processing: uses processing
image_similarity.py - CLI build and query → Model indexes: stores indexes
README.md - Caltech subset preparation → Caltech queries and labels: prepares subset
scripts/prepare_caltech101.py - Retrieval evaluation → Caltech queries and labels: reads queries
evaluate.py - Retrieval evaluation → Feature extractors: embeds images
evaluate.py - Retrieval evaluation → Retrieval processing: processes rankings
evaluate.py - Retrieval evaluation → Ranking metrics: summarizes scores
evaluate.py - Retrieval evaluation → Evaluation outputs: writes reports
README.md - DINOv2 fine-tuning → Caltech queries and labels: trains on database
finetune.py - DINOv2 fine-tuning → Fine-tuned checkpoint: writes weights
finetune.py - Feature extractors → Fine-tuned checkpoint: loads fine-tuned weights
extractors.py