StarNew/flexdepth

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

Overview

FlexDepth is a family of self-supervised monocular depth estimation models designed for robust driving perception. It introduces a Scale-Driven Decoder (SDD) with adaptive component selection, enabling a single architecture to span five model scales — from ultra-lightweight (1.5M params, 0.7 GFLOPs) to high-accuracy (32.3M params, 24.6 GFLOPs).

Key Features

• Five model scales: Nano (N), Small (S), Medium (M), Large (L), X-Large (X)
• Scale-Driven Decoder (SDD): Adaptive component selection based on model scale
• High-Efficiency Bottleneck (HEB): For small models (N, S) — maximizes efficiency
• High-Performance Bottleneck (HPB): For large models (M, L, X) — maximizes accuracy
• Dynamic upsampling: Sharper depth boundaries via learned upsampling
• Two-stage static-dynamic decoupled training: Handles dynamic scenes in driving scenarios
• YOLO11-based encoder: Leverages rich visual representations from YOLO segmentation pretraining

Model Zoo

KITTI-trained Models

Model	Params	GFLOPs	Abs Rel ↓	Sq Rel ↓	RMSE ↓	RMSE log ↓	δ<1.25 ↑	δ<1.25² ↑	δ<1.25³ ↑
Flex-Nano	1.5M	0.7	0.110	0.794	4.678	0.184	0.878	0.961	0.983
Flex-Small	6.1M	2.8	0.104	0.713	4.458	0.179	0.890	0.964	0.983
Flex-Medium	12.7M	10.0	0.096	0.639	4.253	0.172	0.903	0.968	0.985
Flex-Large	15.2M	11.5	0.095	0.642	4.199	0.171	0.906	0.968	0.984
Flex-X-Large	32.3M	24.6	0.093	0.605	4.114	0.167	0.910	0.969	0.985

Cityscapes-trained Models

Model	Params	GFLOPs	Abs Rel ↓	Sq Rel ↓	RMSE ↓	RMSE log ↓	δ<1.25 ↑	δ<1.25² ↑	δ<1.25³ ↑
Flex-Nano	1.5M	0.6	0.107	1.261	6.133	0.164	0.893	0.971	0.989
Flex-Small	6.1M	2.2	0.100	1.078	5.813	0.153	0.904	0.975	0.991
Flex-Medium	12.7M	8.0	0.089	0.885	5.358	0.143	0.917	0.979	0.993
Flex-Large	15.2M	9.2	0.087	0.911	5.310	0.139	0.924	0.981	0.993
Flex-X-Large	32.3M	19.7	0.086	0.877	5.268	0.137	0.926	0.982	0.993

Efficiency

Model	FPS (Snapdragon 8 Elite)	FPS (RTX 2080 Ti)
Flex-Nano	37.6	180+
Flex-Small	18.6	120+
Flex-Medium	5.8	60+
Flex-Large	5.2	55+
Flex-X-Large	3.0	40+

Model Files

Each model consists of two weight files:

├── kitti/
│   ├── flex_n/
│   │   ├── encoder.pth    # YOLO11-based encoder weights
│   │   └── depth.pth      # Scale-Driven Decoder weights
│   ├── flex_s/
│   ├── flex_m/
│   ├── flex_l/
│   └── flex_x/
└── cs/
    ├── flex_n/
    ├── flex_s/
    ├── flex_m/
    ├── flex_l/
    └── flex_x/

Usage

Installation

conda create -n flexdepth python=3.10
conda activate flexdepth

pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Download Weights

from huggingface_hub import snapshot_download

# Download all models
snapshot_download(repo_id="StarNew/flexdepth", local_dir="./models")

# Or download a specific model
from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="StarNew/flexdepth",
    filename="kitti/flex_n/encoder.pth",
    local_dir="./models"
)
hf_hub_download(
    repo_id="StarNew/flexdepth",
    filename="kitti/flex_n/depth.pth",
    local_dir="./models"
)

Evaluation

# Flex-Nano on KITTI
python evaluate_depth.py --png --eval_mono --scale 4 \
    --encoder_model_type yolo11n-seg --decoder_model_type flexn \
    --load_weights_folder ./models/kitti/flex_n \
    --data_path <kitti_data_path> --split_path <splits_path>

# Flex-X-Large on KITTI
python evaluate_depth.py --png --eval_mono --scale 4 \
    --encoder_model_type yolo11x-seg --decoder_model_type flexx \
    --load_weights_folder ./models/kitti/flex_x \
    --data_path <kitti_data_path> --split_path <splits_path>

ONNX Export

python export_onnx.py --encoder_model_type yolo11n-seg --decoder_model_type flexn \
    --load_weights_folder ./models/kitti/flex_n --scales 4 --export_name flex-n

Comparison with Depth Anything V2

On the KITTI Eigen benchmark with dense ground truth and least-squares alignment:

Method	Type	Params	GFLOPs	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	Self-Supervised	32M	25	640×192	0.063	0.952

FlexDepth achieves comparable or better accuracy than Depth Anything V2 with ~13× fewer parameters and ~78× fewer GFLOPs at similar resolution.

Citation

@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.

Our code is built upon Monodepth2, Manydepth, and Ultralytics.