CMU's MidVCL Trains Vision AI Twice as Fast by Mimicking Human Perception
A CVPR paper shows that teaching networks to see shapes and surfaces first, before semantics, cuts contrastive learning time in half.
ProTakeaways
- CMU's CVPR 2025 paper adds a perceptual warm-up before contrastive pretraining, halving ResNet-18 convergence time.
- Shape Prototypical Contrastive Learning clusters figure-ground silhouettes with K-Means and uses centroids as shape prototypes.
- Retinex decomposition provides reflectance and shading views, each specialized to different downstream tasks.
- MidVCL scores 32.1 mIoU on ADE20K and 0.1354 RME on NYU Depth, beating SimCLR, MoCoV2, BYOL and DINO.
- S-PCL is the only method with shape bias above 0.5 on the model-vs-human benchmark, boosting OOD robustness.
- Limitations: tested only on ResNet-18 with ImageNet-100 and STL-10; no official code release yet.
A group from Carnegie Mellon University is challenging a foundational assumption of modern self-supervised vision: that a network should learn to group semantically similar images directly, in one shot. Their CVPR 2025 paper argues that skipping the intermediate perceptual steps human vision performs (edges, surfaces, shapes) is what makes contrastive pretraining slow and texture-biased. Insert those steps as a warm-up, and a ResNet-18 converges twice as fast while ending up with better representations for classification, segmentation, and depth.
Why contrastive learning cheats
Contrastive frameworks like SimCLR and MoCo pull together augmented views of the same image and push apart views from different images. The objective is efficient but blunt. End-to-end training in these networks tends to seize on the simplest available features, bypassing intermediate representations and drifting toward texture bias, since the loss groups semantically similar scenes without first building features sensitive to surfaces and boundaries.
The consequence is a well-known misalignment with human vision. Recent work argues that most discriminative models remain strongly texture-biased and fragile under large style changes, because they chase semantic representations while ignoring intermediate-level ones. The CMU fix borrows David Marr's staged theory of perception, boundary and surface first, semantics later, and injects it directly into pretraining.