# DINOv2

Cell morphology profiling has relied on cell segmentation and handcrafted features, which are annotation-hungry and hard to reuse across laboratories. Researchers pretrained four self-supervised models on 119,876 unannotated compounds and about 1.77 million JUMP–CP images, then froze the weights and extracted image embeddings directly. On 2,464 EU-OPENSCREEN Bioactives compounds, DINO and DINOv2 beat CellProfiler by 31–37% in site-averaged mean average precision for mechanism-of-action retrieval and consistently surpassed masked-based models.

Cell morphology profiling has relied on cell segmentation and handcrafted features, which are annotation-hungry and hard to reuse across laboratories. Researchers pretrained four self-supervised models on 119,876 unannotated compounds and about 1.77 million JUMP–CP images, then froze the weights and extracted image embeddings directly. On 2,464 EU-OPENSCREEN Bioactives compounds, DINO and DINOv2 beat CellProfiler by 31–37% in site-averaged mean average precision for mechanism-of-action retrieval and consistently surpassed masked-based models.

Image compression, restoration and generation all need to measure how different two images look to a person, yet the most accurate perceptual distances are fitted to human judgments and lose accuracy when resolution changes. JLD turns the Jacobian of a frozen vision encoder's output with respect to its early features into one fixed metric tensor, fitted once from 100 unlabeled images in about 35 seconds with no human labels. Across TID2013, CSIQ, LIVE and KADID-10k it reaches a mean Spearman correlation of 0.926, above 0.915 for VSI and 0.900 for DISTS; on TID2013 its lens-term correlation falls only from 0.850 to 0.845 when resolution doubles from 256 to 512 pixels, where DISTS drops from 0.815 to 0.717.

Image compression, restoration and generation all need to measure how different two images look to a person, yet the most accurate perceptual distances are fitted to human judgments and lose accuracy when resolution changes. JLD turns the Jacobian of a frozen vision encoder's output with respect to its early features into one fixed metric tensor, fitted once from 100 unlabeled images in about 35 seconds with no human labels. Across TID2013, CSIQ, LIVE and KADID-10k it reaches a mean Spearman correlation of 0.926, above 0.915 for VSI and 0.900 for DISTS; on TID2013 its lens-term correlation falls only from 0.850 to 0.845 when resolution doubles from 256 to 512 pixels, where DISTS drops from 0.815 to 0.717.

DisParQ learns spatially grounded, discrete concept representations from a powerful frozen vision-only self-supervised backbone without class labels or language supervision: each image patch is assigned to exactly one concept from a learnable prototype dictionary, only a sparse subset of concepts may activate per image, and continuous residuals quantized into discrete attributes capture how each concept varies; a spatial decoder reconstructs the backbone's representation from concepts and attributes alone. Across seven datasets it reaches 83.2% top-1 on ImageNet linear probing, closely matching its frozen DINOv2 teacher, achieves higher concept consistency than language-aligned models, remains competitive on fine-grained recognition, and enables cross-category part-based retrieval.

DisParQ learns spatially grounded, discrete concept representations from a powerful frozen vision-only self-supervised backbone without class labels or language supervision: each image patch is assigned to exactly one concept from a learnable prototype dictionary, only a sparse subset of concepts may activate per image, and continuous residuals quantized into discrete attributes capture how each concept varies; a spatial decoder reconstructs the backbone's representation from concepts and attributes alone. Across seven datasets it reaches 83.2% top-1 on ImageNet linear probing, closely matching its frozen DINOv2 teacher, achieves higher concept consistency than language-aligned models, remains competitive on fine-grained recognition, and enables cross-category part-based retrieval.

This work proposes RIM (Retrieval-In-Matching), which renders UAV-viewpoint references from Google 3D Tiles across locations, altitudes, and orientations, adapts SALAD in two stages with pose-near positives and geographically distant hard negatives and then re-ranks Top-K candidates by local geometric consistency, and further freezes the adapted DINOv2-B retriever while distilling a local-descriptor decoder from its token field and a shallow VGG19 detail stream so that one query-side DINOv2-B forward supports both SALAD retrieval and local description; evaluated zero-shot on the reconstructed EPFL Urbanscape and self-collected Chang'an Park datasets, both geographically disjoint from the training data, RIM outperforms ten retrieval baselines, improving Recall@1 over SALAD by 8.55/13.

This work proposes RIM (Retrieval-In-Matching), which renders UAV-viewpoint references from Google 3D Tiles across locations, altitudes, and orientations, adapts SALAD in two stages with pose-near positives and geographically distant hard negatives and then re-ranks Top-K candidates by local geometric consistency, and further freezes the adapted DINOv2-B retriever while distilling a local-descriptor decoder from its token field and a shallow VGG19 detail stream so that one query-side DINOv2-B forward supports both SALAD retrieval and local description; evaluated zero-shot on the reconstructed EPFL Urbanscape and self-collected Chang'an Park datasets, both geographically disjoint from the training data, RIM outperforms ten retrieval baselines, improving Recall@1 over SALAD by 8.55/13.