Research map / Vision-language and multimodal / Human and embodied modeling

DepthCues: Evaluating Monocular Depth Perception in Large Vision Models

CVPR 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Human perception modeling · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces DepthCues, a probing benchmark of six human monocular depth cues, and uses it to evaluate 20 pre-trained vision models and show that LoRA fine-tuning on the cues improves linear-probe depth estimation.

Problem

It is unclear whether human-like monocular depth cues (elevation, light and shadow, occlusion, perspective, size, texture gradient) emerge in large pre-trained vision models that never see explicit depth supervision.

Approach

Adapt existing datasets for six depth cues into probing tasks, train lightweight MLP or attentive probes on frozen features of 20 models, and test whether depth cue understanding correlates with depth estimation; also fine-tune DINOv2 and CLIP with LoRA on the cue tasks to boost depth linear probing.

Key results

• DepthAnythingv2-b14 achieves the highest average performance on DepthCues, with DINOv2 and Stable Diffusion ranking in the top five.
• Multi-view models (CroCo, LRM, DUSt3R) rank in the top four on texture-gradient, outperforming DINOv2.
• LoRA fine-tuning of DINOv2 on DepthCues with concatenated features raises NYUv2 linear-probe accuracy from 87.78% to 88.46% and lowers DIW WHDR from 11.99 to 11.72.

Measured effects reported in the paper

Surprising findings

• MiDaS, a dedicated depth model, is outperformed on DepthCues by several models pre-trained without depth supervision, such as DeiT, SigLIP, and DINO.
• CLIP obtains the lowest DepthCues performance, while SigLIP trained on much more data performs significantly better.

Limitations

• Analysis uses released pre-trained weights and does not isolate the effect of pre-training objectives or datasets.
• Motion-based depth cues (ego-motion, scene motion) are not covered; focus is on static image models.
• Fine-tuning gains on depth estimation are small and generalization of fine-tuned features alone decreases, requiring concatenation with original features.

Experimental setup

Datasets: HLW SOBA COCOA NaturalScene KITTI SUN-RGBD DTD NYUv2 DIW ImageNet-1k

Benchmarks: DepthCues NYUv2 DIW ImageNet-1k

Metrics: classification accuracy success rate accuracy (delta < 1.25) WHDR top-1 accuracy Spearman rank correlation

Baselines: MiDaS CLIP SigLIP Stable Diffusion 2.1 CroCo DUSt3R LRM end-to-end ResNet18 end-to-end ViT-B/16 trivial coordinate-map baseline

Builds on: DINOv2 LoRA Probing 3D awareness (Zhou et al.)

Closest papers by meaning

• The 3D-PC: a benchmark for visual perspective taking in humans and machines (ICLR 2025 · similarity 0.88)
• Depth Anything V2 (NeurIPS 2024 · similarity 0.88)
• 3D Visual Illusion Depth Estimation (NeurIPS 2025 · similarity 0.87)
• Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.87)
• Depth Prompting for Sensor-Agnostic Depth Estimation (CVPR 2024 · similarity 0.87)
• Seurat: From Moving Points to Depth (CVPR 2025 · similarity 0.87)