🎉🎉🎉

🐮

🐂

🐂

👏

InfiniDepth: Main Results and Key Findings

Overview

InfiniDepth introduces a revolutionary approach to monocular depth estimation by representing depth as neural implicit fields rather than discrete grids. This enables arbitrary-resolution and fine-grained depth prediction, addressing fundamental limitations of existing methods.

Key Innovations and Results

1. Neural Implicit Field Representation

Figure 1 showcases InfiniDepth's three main capabilities:

• (a) Arbitrary-resolution depth estimation - can query depth at any continuous coordinate
• (b) Fine-grained point clouds with geometric detail preservation
• (c) Enhanced novel view synthesis with fewer holes and artifacts

The core insight is modeling depth as a continuous function:

d_I(x,y) = N_θ(I,(x,y))

where any 2D coordinate can be mapped to a depth value, breaking free from grid constraints.

2. Multi-Scale Local Implicit Decoder

Figure 2 illustrates the two-module architecture:

Feature Query (a):

• Extracts multi-scale features from ViT encoder layers
• Constructs feature pyramid with different spatial resolutions
• Uses bilinear interpolation to query features at continuous coordinates

Depth Decoding (b):

• Hierarchically fuses features from high-to-low resolution
• Employs residual gated fusion blocks
• Predicts depth through lightweight MLP head

3. Infinite Depth Query Strategy

Figure 3 demonstrates a key breakthrough: traditional per-pixel depth prediction creates density imbalance due to perspective projection and surface orientation effects. InfiniDepth's adaptive query strategy:

• Computes adaptive weights: w(x,y) = d_I(x,y)² / |n(x,y)·v(x,y)| + ε
• Allocates sub-pixel query budgets proportionally to 3D surface area
• Generates uniformly distributed 3D points on object surfaces

4. High-Quality Internal Geometry

Figure 4 shows that the model learns high-quality internal geometry, with normal maps computed through autograd revealing detailed surface structure.

Quantitative Results

Synthetic Benchmark (Synth4K)

The paper introduces Synth4K, a new 4K synthetic benchmark from five games with diverse scenes and geometric details.

Relative Depth Estimation (Table 1):

• InfiniDepth achieves state-of-the-art performance across all metrics
• Particularly strong in high-frequency (HF) masked regions
• δ₁ accuracy significantly outperforms baselines

Metric Depth Estimation (Table 2):

• Combined with sparse depth inputs ("Ours-Metric")
• Superior performance on stricter thresholds (δ₀.₀₁, δ₀.₀₂, δ₀.₀₄)

Real-World Benchmarks

Relative Depth (Table 3):

• Competitive performance on KITTI, ETH3D, NYUv2, ScanNet, DIODE
• On par with current SOTA methods

Metric Depth (Table 4):

• Clear improvements over existing metric depth methods
• Outperforms Marigold-DC, Omni-DC, PriorDA, PromptDA

Qualitative Comparisons

Depth Map Quality

Figure 5 shows:

• First two rows: Synth4K predictions with superior detail preservation
• Bottom row: Real-world data with low-resolution input
• Highlighted boxes demonstrate fine-detail recovery capabilities

Metric Depth Results

Figure 6 highlights geometric detail recovery in high-frequency regions, showing cleaner edges and better preservation of fine structures.

Novel View Synthesis

Figure 8 demonstrates superior novel view synthesis under large viewpoint shifts:

• InfiniDepth produces complete, stable results
• ADGaussian baseline shows noticeable geometric holes and artifacts
• The infinite depth query strategy ensures uniform point distribution

Key Ablation Studies

Depth Representation Effectiveness (Table 5)

• Neural implicit fields significantly outperform discrete grid representations
• Gains more pronounced in metric depth estimation with sparse inputs

Multi-Scale Feature Query

• Multi-scale mechanism brings substantial improvements
• Single-scale baseline performs considerably worse

Computational Efficiency (Table 6)

• Decoder has the lowest parameter count among compared methods
• Competitive computational efficiency despite superior detail preservation

Impact and Significance

1. Resolution Independence: Breaks free from training resolution constraints
2. Fine Detail Preservation: Excels in geometrically complex regions
3. Multi-Task Versatility: Effective for both relative and metric depth estimation
4. Downstream Applications: Benefits novel view synthesis, 3D reconstruction pipelines
5. Benchmark Contribution: Synth4K enables better evaluation of high-resolution depth estimation

Limitations and Future Work

• No explicit temporal consistency for video applications
• Future work: extend to multi-view settings for improved temporal stability and 3D consistency

InfiniDepth represents a fundamental shift in depth estimation, moving from discrete grid representations to continuous neural implicit fields, enabling unprecedented resolution scalability and detail preservation capabilities.

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

• Long-LRM++: Preserving Fine Details in Feed-Forward Wide-Coverage Reconstruction (2025)
• Depth Anything 3: Recovering the Visual Space from Any Views (2025)
• CloseUpShot: Close-up Novel View Synthesis from Sparse-views via Point-conditioned Diffusion Model. (2025)
• 360-GeoGS: Geometrically Consistent Feed-Forward 3D Gaussian Splatting Reconstruction for 360 Images (2026)
• GVSynergy-Det: Synergistic Gaussian-Voxel Representations for Multi-View 3D Object Detection (2025)
• Re-Depth Anything: Test-Time Depth Refinement via Self-Supervised Re-lighting (2025)
• Depth Any Panoramas: A Foundation Model for Panoramic Depth Estimation (2025)

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend