Research map / Label-efficient and robust learning / Self-supervised and contrastive
PPEA-Depth: Progressive Parameter-Efficient Adaptation for Self-Supervised 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 PPEA-Depth, a two-stage progressive parameter-efficient adaptation (encoder and decoder adapters on a frozen RepLKNet) that transfers ImageNet-pretrained models to self-supervised monocular depth, first on KITTI (static) then on CityScapes/DDAD (dynamic).
Problem
Self-supervised monocular depth estimation relies on a static-scene assumption, so performance degrades in dynamic scenes, and full fine-tuning of pre-trained models on small depth datasets can cause catastrophic forgetting of pre-trained knowledge.
Approach
Freeze the pre-trained RepLKNet encoder and train lightweight encoder adapters (3x3 conv bottleneck in RepLKBlocks, linear in ConvFFNs) plus a decoder adapter in a static-scene stage, then in a dynamic-scene stage freeze the whole U-Net and tune only the adapters, so depth priors learned earlier are preserved.
Key results
- On KITTI Eigen split, PPEA-Depth (RepLKNet-L, 2 frames) reaches AbsRel 0.088 and delta<1.25 0.917, with the authors reporting state-of-the-art results on most metrics.
- On CityScapes, the scene-adaptation stage with adapters reaches AbsRel 0.100 and delta<1.25 0.904 versus 0.130 AbsRel for training from scratch.
- On DDAD, PPEA-Depth (L) with a single frame achieves AbsRel 0.130 and delta<1.25 0.846, outperforming the compared baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add encoder adapters instead of full encoder fine-tuning | RepLKNet-B domain adaptation · KITTI | AbsRel: 0.092 -> 0.092 | No clear effect |
| increase encoder-adapter bottleneck ratio from 0.0625 to 0.25 (ablation) | RepLKNet-B domain adaptation · KITTI | AbsRel: 0.092 -> 0.090 | Helped |
| initialize scene adaptation from KITTI domain-adaptation weights | PPEA-Depth · CityScapes | AbsRel: 0.130 -> 0.100 | Helped |
| use decoder adapter instead of full decoder fine-tuning | PPEA-Depth initialized from domain adaptation · CityScapes | AbsRel: 0.103 -> 0.100 | Helped |
| exclude dynamic-object regions from the training loss (ablation) | full fine-tuning · CityScapes | AbsRel: 0.116 -> 0.106 | Helped |
| exclude dynamic-object regions from the training loss (ablation) | adapter tuning · CityScapes | AbsRel: 0.100 -> 0.102 | Hurt |
| add adapters to RepLKBlock rather than ConvFFN only (ablation) | RepLKNet-B encoder · KITTI | AbsRel: 0.100 -> 0.093 | Helped |
| use a convolutional down projector instead of a linear projector (ablation) | encoder adapters · KITTI | AbsRel: 0.095 -> 0.090 | Helped |
| use a convolutional up projector instead of a linear projector (ablation) | encoder adapters · KITTI | AbsRel: 0.090 -> 0.090 | No clear effect |
| add prefix BatchNorm before the encoder adapter (ablation) | encoder adapters · KITTI | AbsRel: 0.095 -> 0.090 | Helped |
| share prefix BatchNorm with the pre-trained block instead of training extra adapter BatchNorm (ablation) | encoder adapters · KITTI | AbsRel: 0.090 -> 0.091 | No clear effect |
| add feature scale 0 alongside scale 3 to decoder-adapter input (ablation) | PPEA-Depth scene adaptation · CityScapes | AbsRel: 0.103 -> 0.100 | Helped |
Surprising findings
- Tuning encoder and decoder adapters with only about 3% of CityScapes training data outperforms full fine-tuning and training from scratch on the full set.
- Tuning the whole encoder plus encoder adapters performed no better than freezing the encoder and tuning adapters only (0.093 vs 0.090 AbsRel on KITTI).
Limitations
- Cost-volume student network requires a fixed depth range, so DDAD (0-200m) evaluation is restricted to the single-frame teacher network.
- Stage 2 adapters must be tuned and stored separately for each target dataset.
- Evaluation only on KITTI, CityScapes and DDAD with a single backbone family (RepLKNet).
Experimental setup
Datasets: KITTI CityScapes DDAD ImageNet
Benchmarks: KITTI CityScapes DDAD
Metrics: AbsRel SqRel RMSE RMSE_log delta<1.25 delta<1.25^2 delta<1.25^3
Baselines: ManyDepth Frozen encoder Full fine-tuning Monodepth2 Scratch training on CityScapes
Builds on: ManyDepth RepLKNet adapter tuning
Closest papers by meaning
- From-Ground-To-Objects: Coarse-to-Fine Self-supervised Monocular Depth Estimation of Dynamic Objects with Ground Contact Prior (CVPR 2024 · similarity 0.88)
- Mining Supervision for Dynamic Regions in Self-Supervised Monocular Depth Estimation (CVPR 2024 · similarity 0.87)
- SQLdepth: Generalizable Self-Supervised Fine-Structured 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 0.86)
- Synthetic-to-Real Self-supervised Robust Depth Estimation via Learning with Motion and Structure Priors (CVPR 2025 · similarity 0.86)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · 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.