FlexDepth: A Flexible Scale-Driven Family for Self-Supervised Monocular Depth Estimation

TL;DR: FlexDepth is a scale-driven family of self-supervised monocular depth models that achieves SOTA across arbitrary scales with minimal computational overhead. Our smallest model, Flex-Nano, requires only 0.7 GFLOPs and achieves 37.6 FPS on mobile platforms.

Accuracy vs. Computational Cost on KITTI

Accuracy (δ<1.25) vs. Computational Cost on KITTI Eigen split. FlexDepth achieves the best trade-off across all model scales.

Abstract

Self-Supervised Monocular Depth Estimation (MDE) has garnered attention in recent years due to its independence from ground truth. However, most existing models are limited to a single scale and exhibit considerable performance degradation in complex driving environments. Networks specifically designed to handle dynamic traffic participants tend to be overly complex, hindering their deployment on resource-constrained automotive edge devices. To address these limitations and move towards robust driving perception, we propose FlexDepth, a scale-driven and flexible family of self-supervised MDE models tailored for challenging road scenarios. FlexDepth employs a two-stage static-dynamic decoupled training strategy, enabling the independent assessment of confidence for both static backgrounds and dynamic road objects. Furthermore, it introduces a meticulously designed Scale-Driven Decoder (SDD) to dynamically select components based on scale size, facilitating efficient feature fusion and the output of high-precision depth maps. Extensive experiments on standard driving benchmarks demonstrate that without any auxiliary information, our model achieves state-of-the-art performance across arbitrary scales with minimal computational overhead. Our smallest model, Flex-Nano, requires only 0.7 GFLOPs and achieves 37.6 FPS on mobile platforms, ensuring reliable real-time perception while maintaining excellent zero-shot generalization.

Method

Overview of the FlexDepth architecture. The training process is structured into two stages: Stage 1 jointly optimizes the Depth Network and PoseNet; Stage 2 introduces the static-dynamic decoupled mask for dynamic scene refinement.

Scale-Driven Decoder

A novel decoder that adaptively selects its constituent components (Post-Convolution, Upsampling, Prediction Head) based on model scale, enabling efficient feature fusion and high-precision depth output across Nano to X-Large variants.

Static-Dynamic Decoupled Training

A two-stage training strategy that independently assesses confidence for static backgrounds and dynamic objects via an adaptive thresholding mechanism, achieving superior robustness in dynamic driving scenes without extra inference overhead.

Flexible Model Family

Five model scales (Nano, Small, Medium, Large, X-Large) covering 0.7 to 24.6 GFLOPs. Flex-Nano achieves 37.6 FPS on mobile platforms, while Flex-X-Large sets new SOTA on KITTI and Cityscapes benchmarks.

Post-Convolution Design

We remove the pre-convolution stage entirely and redesign the post-convolution with two Bottleneck variants: High-Efficiency Bottleneck (HEB) for small models prioritizing gradient flow richness, and High-Performance Bottleneck (HPB) for large models capturing finer-grained local features via CSP structure.

Dynamic Upsampling

Instead of content-agnostic bilinear interpolation that causes blurred boundaries, we predict upsampling offsets adaptively based on semantic content. This enables non-uniform upsampling that generates sharper and more accurate structural boundaries.

Adaptive Masking with MuPredictor

The static-dynamic decoupled mask employs an adaptive thresholding function that dynamically adjusts sensitivity based on feature evolution between early and late training stages, ensuring only stable static pixels contribute to confidence evaluation.

Results

KITTI Eigen Split

Method	Year	Abs Rel ↓	Sq Rel ↓	RMSE ↓	RMSE log ↓	δ<1.25 ↑	δ<1.25² ↑	δ<1.25³ ↑	FLOPs(G) ↓
Lightweight (<10 GFLOPs)
Monodepth2	2019	0.115	0.903	4.863	0.193	0.877	0.959	0.981	8.0
Lite-Mono-Tiny	2023	0.110	0.837	4.710	0.187	0.880	0.960	0.982	2.8
Lite-Mono-Small	2023	0.110	0.802	4.671	0.186	0.879	0.961	0.982	4.7
Lite-Mono	2023	0.107	0.765	4.561	0.183	0.886	0.963	0.983	5.0
PuriLight	2026	0.106	0.747	4.536	0.182	0.890	0.964	0.984	7.1
Flex-Nano (Ours)	2026	0.110	0.794	4.678	0.184	0.878	0.961	0.983	0.7
Flex-Small (Ours)	2026	0.104	0.713	4.458	0.179	0.890	0.964	0.983	2.8
Medium (10–20 GFLOPs)
Lite-Mono-8M	2023	0.101	0.729	4.454	0.178	0.897	0.965	0.983	11.2
GeoDepth	2025	0.100	0.694	4.381	0.176	0.897	0.966	0.984	11.9
TinyDepth	2024	0.096	0.665	4.249	0.171	0.904	0.968	0.985	13.3
Flex-Medium (Ours)	2026	0.096	0.639	4.253	0.172	0.903	0.968	0.985	10.0
Flex-Large (Ours)	2026	0.095	0.642	4.199	0.171	0.906	0.968	0.984	11.5
Large (>20 GFLOPs)
MonoViT	2022	0.099	0.708	4.372	0.175	0.900	0.967	0.984	59.7
FGTO^†	2024	0.096	0.696	4.327	0.174	0.904	0.968	0.985	59.7
DSI-MonoViT	2025	0.096	0.711	4.321	0.172	0.905	0.968	0.984	59.7
Self-Distillation	2025	0.095	0.619	4.156	0.169	0.905	0.969	0.985	64.5
Flex-X-Large (Ours)	2026	0.093	0.605	4.114	0.167	0.910	0.969	0.985	24.6

Quantitative results on KITTI Eigen split (640×192). Flex-X-Large achieves SOTA with only 24.6G FLOPs — 2.4× fewer than the next best method.

Efficiency & Speed

Model	Params(M)	FLOPs(G)	RTX 4090	V100	SD 8 Elite
Flex-Nano	1.5	0.7	1583	774	37.6
Flex-Small	6.1	2.8	1215	606	18.6
Flex-Medium	12.7	10.0	510	231	5.8
Flex-Large	15.2	11.5	467	216	5.2
Flex-X-Large	32.3	24.6	321	130	3.0

Inference speed across platforms. Flex-Nano achieves real-time performance even on mobile devices (Snapdragon 8 Elite).

Qualitative Results

Cityscapes: FlexDepth produces sharper boundaries and more robust depth in dynamic regions compared to DSI and ProDepth.

KITTI: FlexDepth preserves fine-grained details and object contours, outperforming Monodepth2, MonoViT, and Lite-Mono.

Extended Cityscapes qualitative results showing consistent superiority across diverse dynamic scenarios.

Extended KITTI qualitative results demonstrating robust depth estimation across various driving scenes.

Comparison with Foundation Depth Models (Zero-Shot)

Method	Type	Params	FLOPs(G)	Resolution	Abs Rel ↓	δ<1.25 ↑
DA2 (ViT-L)	Zero-Shot	335M	1947	1722×518	0.070	0.956
DA2 (ViT-S)	Zero-Shot	25M	137	1722×518	0.077	0.944
DA2 (ViT-L)	Zero-Shot	335M	276	644×196	0.092	0.915
DA2 (ViT-S)	Zero-Shot	25M	19	644×196	0.110	0.881
Flex-X-Large (Ours)	Self-Supervised	32M	25	640×192	0.063	0.952

Comparison with Depth Anything V2 on KITTI Eigen split. Flex-X-Large achieves superior Abs Rel (0.063 vs. 0.070) with 78× fewer FLOPs and 10× fewer parameters than DA2 ViT-L at highest resolution.

Zoomed Comparison with Depth Anything V2 (Zero-Shot)

Click on a scene, then click the original image or any depth map to view the zoomed comparison in fullscreen. DA2 results are from its zero-shot setting; Flex-X-Large is self-supervised. Red boxes indicate magnified regions.

DA2 ViT-L (1722×518)

DA2 ViT-L (644×196)

DA2 ViT-S (1722×518)

DA2 ViT-S (644×196)

Flex-X-Large (Ours)

Video Comparison

FlexDepth on KITTI

Depth estimation results of all FlexDepth model variants on the KITTI test dataset, demonstrating consistent depth quality from Flex-Nano to Flex-X-Large.

FlexDepth on Cityscapes

Depth estimation results on the Cityscapes test dataset. FlexDepth produces sharp object boundaries and robust depth predictions even in highly dynamic urban scenarios.

Model Comparison — KITTI Drive 0005

Comparison with 6 baselines (LiteMono-Tiny, LiteMono-Small, LiteMono, LiteMono-8M, MonoDepth2, MonoViT) and 5 FlexDepth variants. Layout: Original Image | Baselines | Ours.

Model Comparison — KITTI Drive 0009

Additional driving sequence comparison on KITTI. FlexDepth models demonstrate superior depth estimation quality across all model scales with sharper boundaries.

Flex-Nano on Mobile (Snapdragon 8 Elite)

Real-time inference of Flex-Nano (ONNX, 0.7 GFLOPs) on a Snapdragon 8 Elite device, achieving 37.6 FPS. Recorded October 2025.

Flex-X-Large on Mobile (Snapdragon 8 Elite)

Flex-X-Large (ONNX, 24.6 GFLOPs) running on a Snapdragon 8 Elite device. The average FPS shown is slightly lower than the 3.0 FPS reported in the paper due to device temperature and OS changes at the time of this recording (July 2026).

Citation

BibTeX

@misc{zhu2026robustdrivingperceptionflexible,
  title={Towards Robust Driving Perception: A Flexible Scale-Driven Family for Self-Supervised Monocular Depth Estimation},
  author={Zhaowen Zhu and Li Zhang and Yujie Chen and Tian Zhang and Yingjie Wang and Mingxia Zhan},
  year={2026},
  eprint={2607.00736},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2607.00736}
}

Acknowledgment

This work is supported by the National Natural Science Foundation of China under Grant 62332016.

This website is built based on the Depth Anything 3 project page template. We thank the DA3 authors for making their template available.

FlexDepth

Towards Robust Driving Perception:A Flexible Scale-Driven Family for Self-Supervised Monocular Depth Estimation