Research map / 3D vision / Depth and stereo

Boosting Self-Supervision for Single-View Scene Completion via Knowledge Distillation

CVPR 2024 · 3D vision › Depth and stereo · approach: Dense depth prediction and completion · 3 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.

Proposes MVBTS, a self-supervised multi-view density-field model, and distills its fused predictions into a single-view BTS-based model (KDBTS) for improved occluded-region occupancy prediction on KITTI-360.

Problem

Single-view scene completion from images is limited because occluded geometry is hard to infer; without accurate 3D ground truth, single-view density-field models underperform.

Approach

Extend BTS density-field prediction to fuse multiple posed views with confidence-weighted aggregation trained self-supervised via volumetric rendering, then use its density predictions as pseudo ground truth via stop-gradient L1 knowledge distillation to supervise a single-view BTS model.

Key results

• KDBTS improves over BTS on KITTI-360 single-view occupancy, e.g. IE accuracy 78.30% vs 77.04% and IE recall 44.00% vs 43.99% (recall roughly unchanged), with O recall 84.78% vs 84.24%.
• MVBTS in the multi-view setting reaches 94.91% O accuracy, 61.73% O precision and 45.23% IE recall, matching or beating IBRnet on most metrics.
• On KITTI monocular depth, KDBTS (Abs Rel 0.105) performs close to BTS (0.102) and is outperformed by DevNet (0.095), which is trained explicitly for depth.

Measured effects reported in the paper

Surprising findings

• Attention layers, adopted from GeoNeRF, hinder MVBTS performance rather than helping.
• KDBTS, with a smaller decoder, performs on par with the larger multi-view MVBTS model in single-view occupancy.

Limitations

• Static scene assumption; moving objects cause drawn-out shadows and conflicting information.
• MVBTS requires multiple views and relative poses at inference, adding computational overhead.
• Depth estimation performance is close to but below methods trained explicitly for depth.

Experimental setup

Datasets: KITTI KITTI-360

Benchmarks: KITTI-360 occupancy prediction KITTI monocular depth evaluation

Metrics: occupancy accuracy occupancy precision occupancy recall invisible-empty accuracy invisible-empty precision invisible-empty recall Abs Rel RMSE alpha<1.25

Baselines: BTS PixelNeRF IBRnet MonoDepth2 DevNet

Builds on: Behind the Scenes (BTS) NeRF IBRnet MonoDepth2

Closest papers by meaning

• DistillNeRF: Perceiving 3D Scenes from Single-Glance Images by Distilling Neural Fields and Foundation Model Features (NeurIPS 2024 · similarity 0.88)
• Know Your Neighbors: Improving Single-View Reconstruction via Spatial Vision-Language Reasoning (CVPR 2024 · similarity 0.87)
• Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · similarity 0.87)
• Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation (CVPR 2024 · similarity 0.86)
• A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.86)
• Zero-Shot Novel View and Depth Synthesis with Multi-View Geometric Diffusion (CVPR 2025 · similarity 0.86)