Siamese Coding Network for Near-Duplicate Image Detection

The project structure is:

• /datasets - should be downloaded from appropriate sources (CIFAR10, Imagenette, UKBench, Imagenet, etc.)

• /models - computed model weights are stored here and reused if possible

• /data - is where all computed vectors (embeddings) are stored

• /logs - TendorBoard logs and projections

• /nnfaiss - faiss analysis code

• /sidd - Siamese models and training code

• /scripts - extra scripts for data preparation

• /siftbof - scripts for the baseline method SIFT-BoF

• batch.sh - SLURM batch script for running experiments

• runner*.py - SLURM entry points for experiments

All the experiments are peformed using Python 3.10 and Tensorflow 2.10+.
On a SLURM cluster provided by L3S Research Center utilizing A100 Nvidia GPUs.

An interactive demo with real-time duplicate detection is Work In Progress.

CNN Models & Weights

All the models are stored in the model directory. Most of them, except Alexnet have Imagenet1k pre-trained weights. Alexnet was trained on Imagenet1k using our script.