نمونه کار Image Similarity Search meta faiss

• Built a fine grained image retrieval / metric learning system for a car accessories dataset (413 classes), learning 512 D embeddings for visual similarity search and ranking. • Uses ConvNeXtV2 (tiny/base) backbones with multi scale feature extraction and a learned embedding head to output compact retrieval vectors. • Applies Pairwise Cross Attention (PWCA) for distractor aware features and GeM pooling to strengthen salient regions before embedding projection. • Trains embeddings with Proxy Anchor loss using learnable class proxies and cosine similarity. • Uses a 3 stage curriculum (224 320 384 px) with differential LRs, gradient accumulation, and increasing augmentation strength to stabilize fine tuning • Evaluates retrieval quality with Recall@1/5/10 and mAP, using cosine similarity matrices of embeddings