IC-281SLIP's intra-modal self-supervised loss reduces intra-modal misalignment, making inter-modal inversion unnecessary for image retrieval

Marco Mistretta, Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, Andrew D. Bagdanov

SourceCross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion

When the same image-to-image retrieval experiment is performed on SLIP (which adds a SimCLR-based intra-modal loss to the CLIP contrastive loss during pre-training), the OTI-inverted features achieve nearly identical performance to native image features (e.g., B/16: 35.1 vs 35.3 average MAP), in stark contrast to CLIP where OTI yields a 2-3% boost. This demonstrates that the intra-modal loss during pre-training effectively mitigates the intra-modal misalignment, and that the misalignment observed in CLIP is a direct consequence of the absence of intra-modal constraints in the training objective.

Evidence
correlational
Key metric
SLIP B/16: intra-modal 35.1, inter-modal (OTI) 35.3 (Δ=0.2). SLIP L/16: intra-modal 36.7, inter-modal (OTI) 37.1 (Δ=0.4). Compare CLIP B/32: 39.2 vs 41.2 (Δ=2.0), L/14: 52.0 vs 55.3 (Δ=3.3).
Caveat
SLIP uses ViT-B/16 and ViT-L/16 backbones while the CLIP comparison uses ViT-B/32 and ViT-L/14, so the comparison is not perfectly controlled for backbone size.
Model
SLIP, CLIP / CLIP-ViT (LC)
Concepts
Distance preservation
Datasets
CUB-200-2011 / CUB200 [eval], Stanford Online Products [eval], ROxford [eval], Stanford Cars [eval], Oxford Pets [eval], Flowers-102 [eval], FGVC Aircraft [eval], DTD [eval], EuroSAT [eval], Food-101 [eval], SUN397 [eval], Caltech-101 [eval], UCF101 [eval], ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval]
Methods
OTI [primary]
Related findings
IC-280
Extraction
automatic-extraction