Research map / 3D vision / Depth and stereo
DCDepth: Progressive Monocular Depth Estimation in Discrete Cosine Domain
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 DCDepth, which predicts DCT frequency coefficients of 8x8 depth patches progressively from low to high frequency with a GRU-based head, achieving SoTA monocular depth on NYU-Depth-V2, TOFDC, and KITTI.
Problem
Monocular depth estimation typically predicts depth per pixel in the spatial domain, neglecting local inter-pixel correlations and relying on a single forward pass.
Approach
Transform depth patches into the discrete cosine domain and predict DCT coefficients progressively from low to high frequency using a GRU-based Progressive Prediction Head, with a DCT-based downsampling inside a Pyramid Feature Fusion module.
Key results
- On NYU-Depth-V2, DCDepth reaches Abs Rel 0.085 and delta<1.25 0.940, the best among compared methods including VA-DepthNet (0.086, 0.937).
- On TOFDC, Abs Rel improves to 0.188 from VA-DepthNet's 0.234 (about 19.7%) and RMSE to 0.565 from 0.619.
- On KITTI official split, iRMSE 10.12 is best versus 10.62 for NDDepth; on the Eigen split, RMSE 2.044 is best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| progressively estimate frequency coefficients using the PPH rather than direct convolutional prediction with PixelShuffle (ablation) | DCDepth on NYU-Depth-V2 · NYU-Depth-V2 test | RMSE: 0.318 -> 0.304 | Helped |
| predict frequency-domain depth with inverse DCT rather than spatial-domain depth with convolution and PixelShuffle (ablation) | DCDepth on NYU-Depth-V2 · NYU-Depth-V2 test | RMSE: 0.318 -> 0.315 | Helped |
| add the PPH progressive prediction scheme (ablation) | DCDepth on NYU-Depth-V2 · NYU-Depth-V2 test | RMSE: 0.315 -> 0.304 | Helped |
| aggregate multi-scale features with PFF instead of processing only the last-scale feature with convolution (ablation) | DCDepth on NYU-Depth-V2 · NYU-Depth-V2 test | RMSE: 0.309 -> 0.304 | Helped |
| replace DCT-based downsampling in PFF with bilinear downsampling (ablation) | DCDepth on NYU-Depth-V2 · NYU-Depth-V2 test | RMSE: 0.304 -> 0.305 | Hurt |
| replace DCT-based downsampling in PFF with PixelUnshuffle (ablation) | DCDepth on NYU-Depth-V2 · NYU-Depth-V2 test | RMSE: 0.304 -> 0.306 | Hurt |
| increase iterative prediction steps from 1 to 9 (ablation) | DCDepth on NYU-Depth-V2 · NYU-Depth-V2 test | RMSE: 0.310 -> 0.304 | Helped |
| increase iterative prediction steps from 1 to 9 (ablation) | DCDepth on NYU-Depth-V2 · NYU-Depth-V2 test | Speed (FPS): 31.55 -> 14.24 | Hurt |
| use DCDepth rather than VA-DepthNet | Swin-Large monocular depth estimation · TOFDC test | Abs Rel: 0.234 -> 0.188 | Helped |
| use DCDepth rather than VA-DepthNet | Swin-Large monocular depth estimation · KITTI Eigen split | RMSE: 2.093 -> 2.044 | Helped |
| use DCDepth rather than NDDepth | Swin-Large monocular depth estimation · KITTI official split | iRMSE: 10.62 -> 10.12 | Helped |
Surprising findings
- Using 1 progressive step gives 31.55 FPS with only a small drop in accuracy, while 9 steps reduce speed to 14.24 FPS.
Limitations
- Sparse ground truth (e.g., KITTI LiDAR) gives inefficient supervision of frequency coefficients; method works better with dense supervision.
- Performance depends on dense supervision; sparse-GT training strategies are left to future work.
Experimental setup
Datasets: NYU-Depth-V2 TOFDC KITTI
Benchmarks: NYU-Depth-V2 TOFDC KITTI Eigen split KITTI official split
Metrics: Abs Rel Sq Rel RMSE log10 RMSE log iRMSE delta<1.25 delta<1.25^2 delta<1.25^3 SILog
Baselines: DORN VNL BTS AdaBins LocalBins NeWCRFS BinsFormer PixelFormer IEBins MG-Depth NDDepth VA-DepthNet
Builds on: NeWCRFS PixelFormer IEBins Swin Transformer CRF decoder
Closest papers by meaning
- GeoDepth: From Point-to-Depth to Plane-to-Depth Modeling for Self-Supervised Monocular Depth Estimation (CVPR 2025 · similarity 0.88)
- SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation (CVPR 2025 · similarity 0.87)
- UniDepth: Universal Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.87)
- MonSter: Marry Monodepth to Stereo Unleashes Power (CVPR 2025 · similarity 0.87)
- BetterDepth: Plug-and-Play Diffusion Refiner for Zero-Shot Monocular Depth Estimation (NeurIPS 2024 · similarity 0.87)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.87)
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.