Research map / 3D vision / Depth and stereo
SQLdepth: Generalizable Self-Supervised Fine-Structured Monocular Depth Estimation
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 SQLdepth, a self-supervised monocular depth estimator that builds a self-cost volume from dot products between pixel features and coarse-grained transformer queries, then decodes depth via counting-based depth bins and probabilistic combination.
Problem
Self-supervised monocular depth estimation from video often fails to recover fine-grained scene structure and generalizes poorly, since depth is inferred from immediate visual features.
Approach
A Self Query Layer computes a self-cost volume of dot products between each pixel embedding and a set of coarse-grained queries generated at runtime by a mini-transformer over patch embeddings; depth bins are estimated by counting via pixel-wise softmax-weighted sums, and final depth is a probabilistic combination of bin centers.
Key results
- On KITTI Eigen at 1024x320 with MS training, SQLdepth (ResNet-50) reaches AbsRel 0.082 vs 0.087 for ManyDepth (which uses two frames at test time).
- On Cityscapes, the KITTI-to-Cityscapes fine-tuned model reaches AbsRel 0.106, a 4.4% reduction over the previous best InstaDM.
- Ablation: removing the SQL layer raises AbsRel from 0.091 to 0.114 (17.54% degradation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SQLdepth with ResNet-50 at 1024 x 320 rather than ManyDepth with ResNet-50 | KITTI Eigen · KITTI Eigen, 1024 x 320 | AbsRel: 0.087 -> 0.082 | Helped |
| Use SQLdepth rather than ManyDepth | KITTI improved ground truth · KITTI improved ground truth, 1024 x 320 | AbsRel: 0.055 -> 0.052 | Helped |
| Fine-tune the KITTI-pretrained model on Cityscapes | SQLdepth · Cityscapes, 416 x 128 | AbsRel: 0.125 -> 0.106 | Helped |
| Add the SQL layer with runtime-generated coarse-grained queries (ablation) | SQLdepth encoder-decoder · KITTI Eigen | AbsRel: 0.114 -> 0.091 | Helped |
| Replace learned static queries with runtime-generated coarse-grained queries (ablation) | SQLdepth · KITTI Eigen | AbsRel: 0.102 -> 0.091 | Helped |
| Use coarse-grained rather than fine-grained runtime queries (ablation) | SQLdepth · KITTI Eigen | AbsRel: 0.094 -> 0.091 | Helped |
| Use depth counting approximation rather than fixed uniform bins (ablation) | SQLdepth · KITTI Eigen | AbsRel: 0.114 -> 0.087 | Helped |
| Use depth counting approximation rather than direct bin regression (ablation) | SQLdepth · KITTI Eigen | AbsRel: 0.112 -> 0.087 | Helped |
| Use probabilistic combination rather than global average pooling (ablation) | SQLdepth · KITTI Eigen | AbsRel: 0.115 -> 0.087 | Helped |
| Use probabilistic combination rather than Conv1×1 combination (ablation) | SQLdepth · KITTI Eigen | AbsRel: 0.106 -> 0.087 | Helped |
Surprising findings
- Self-supervised pretraining followed by metric fine-tuning with a ConvNeXt-L backbone reaches AbsRel 0.043 on KITTI, surpassing supervised methods in the appendix comparison.
Limitations
- Cityscapes results depend on fine-tuning from KITTI; the paper highlights that motion masks used by baselines are not used by SQLdepth, which is a comparison-setting caveat.
- Compares mainly against self-supervised baselines with median scaling; the strongest supervised-comparison numbers require self-supervised pretraining plus metric fine-tuning.
Experimental setup
Datasets: KITTI Cityscapes Make3D ImageNet
Benchmarks: KITTI Eigen benchmark KITTI improved ground truth Cityscapes Make3D
Metrics: AbsRel SqRel RMSE RMSE log delta<1.25 delta<1.25^2 delta<1.25^3
Baselines: Monodepth2 ManyDepth EPCDepth CADepth-Net DynamicDepth DIFFNet Depth Hints InstaDM PackNet-SfM HR-Depth
Builds on: Monodepth2 AdaBins ManyDepth SfMLearner
Closest papers by meaning
- GeoDepth: From Point-to-Depth to Plane-to-Depth Modeling for Self-Supervised Monocular Depth Estimation (CVPR 2025 · similarity 0.91)
- From-Ground-To-Objects: Coarse-to-Fine Self-supervised Monocular Depth Estimation of Dynamic Objects with Ground Contact Prior (CVPR 2024 · similarity 0.87)
- UniDepth: Universal Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.87)
- PPEA-Depth: Progressive Parameter-Efficient Adaptation for Self-Supervised Monocular Depth Estimation (AAAI 2024 · similarity 0.87)
- Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks (CVPR 2025 · similarity 0.87)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.