Sovereign 8s.is mirror of standardgalactic/The-Long-Range-Depth-Estimation
  • Python 99.5%
  • Shell 0.5%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-01-10 21:32:15 +08:00
configs Update swinl_22k_w7_nyu.py 2025-06-05 13:28:49 +08:00
depth Update densedepth_head24.py 2025-06-05 14:05:00 +08:00
requirements Add files via upload 2025-06-03 17:15:26 +08:00
splits Add files via upload 2025-06-03 17:15:26 +08:00
tools Add files via upload 2025-06-03 17:15:26 +08:00
LICENSE Add files via upload 2025-05-29 20:52:44 +08:00
README.md Update README.md 2026-01-10 21:18:04 +08:00
script.sh Add files via upload 2025-06-02 14:51:57 +08:00
setup.py Add files via upload 2025-06-02 14:51:57 +08:00

The-Long-Range-Depth-Estimation

This project is a derivative work based on Monocular-Depth-Estimation-Toolbox, which is licensed under the Apache License 2.0.

📖 Introduction

Monocular depth estimation, a cost-effective solution for reconstructing 3D structures from single 2D images, finds widespread applications in robotic navigation, UAV control, and intelligent manufacturing.

Despite significant advancements in deep neural networks, accurately estimating depth for distant objects remains challenging due to insufficient global context modeling.

To address this, we introduce a novel monocular depth estimation framework featuring:

  • **novel heterogeneous hybrid encoder that integrates the initial convolutions of ResNet-50 with the hierarchical self-attention mechanism of Swin-Transformer. This approach balances local detail modeling with the capture of long-range context dependencies, effectively enhancing the expression of depth features in complex environments and distant regions.
  • A Cross-dimensional Feature Fusion (CSF) Module for enhanced global depth understanding.
  • In the decoder, a Depthwise Separable Upsampling Block (DSUB) and Multi-Scale Self-Attention Module (MSA) to refine upsampling and recover detailed spatial information.

Experimental results demonstrate superior long-range prediction accuracy on KITTI and SUN RGB-D datasets.


📊 Method Overview

📌 Architecture Overview

📷 [总体架构72

]

Illustration: Dual-branch Encoder (CNN + Transformer), CSF Fusion Module, DSUB & MSA-based Decoder


📊 Benchmark Performance

Dataset Abs-Rel 🔽 RMSE 🔽 Notes
KITTI 0.050 2.111 checkpoints
NYU-Depth 0.096 0.329 checkpoints
SUN RGB-D 0.142 0.433 Using NYU pre-trained weights, shows strong generalization

📊 Qualitative Results

📷 [image ]


📦 Installation

📋 Prerequisites

  • Linux or macOS (Windows experimental support)
  • Python 3.6+
  • PyTorch 1.3+
  • CUDA 9.2+ (or CUDA 9.0 if building from source)
  • GCC 5+
  • MMCV

🛠️ Installation Steps

Tested with:

  • PyTorch 1.8.0
  • CUDA 11.1
  • Python 3.8
  • Ubuntu 20.04

If your environment is similar, follow these steps:

conda create -n MDE python=3.8
conda activate MDE

conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge

pip install mmcv-full==1.3.13 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.8.0/index.html

git clone https://github.com/sarielbit/The-Long-Range-Depth-Estimation.git
cd The-Long-Range-Depth-Estimation
pip install -e .

pip install future tensorboard

📂 Dataset Preparation

🔹 KITTI

From the official website, download the KITTI dataset. Then unzip the files into data/kitti. Finally, copy split files (whose names start with kitti) in the splits folder into data/kitti.

🔹 NYU Depth v2

Download from: Google Drive Link

Then, download the standard test set from the official site: NYU Depth v2 Dataset

⚠️ The downloaded file will unzip into train and test folders.

🔹 SUN RGB-D

Download from: SUNRGBD Official Download

After downloading, place the dataset under data/SUNRGBD/.

🔹 Example Folder Structure:

The-Long-Range-Depth-Estimation
├── depth
├── tools
├── configs
├── splits
├── data
│   ├── kitti
│   │   ├── input
│   │   │   ├── 2011_09_26
│   │   │   ├── ...
│   │   ├── gt_depth
│   │   │   ├── 2011_09_26_drive_0001_sync
│   │   │   ├── ...
│   │   ├── split_file.txt
│   ├── nyu
│   │   ├── basement_0001a
│   │   ├── ...
│   │   ├── split_file.txt
│   ├── SUNRGBD
│   │   ├── SUNRGBD
│   │   │   ├── kv1
│   │   │   ├── kv2
│   │   │   ├── ...
│   │   ├── split_file.txt

🚀 Usage

🏋️ Training

Single / Multi-GPU Training:

bash tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM} [optional arguments]

Optional Arguments:

  • --no-validate
  • --work-dir ${WORK_DIR}
  • --resume-from ${CHECKPOINT_FILE}
  • --load-from ${CHECKPOINT_FILE}
  • --deterministic

Example:

bash ./tools/dist_train.sh configs/long_rangedepth/swinl_w7_22k_kitti.py 3 --work-dir work_dirs/saves/long_rangedepth/swinl_w7_22k_kitti

🦪 Inference & Testing

Single-GPU Testing:

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] [--eval ${EVAL_METRICS}] [--show]

Example:

python tools/test.py configs/long_rangedepth/swinl_22k_w7_kitti.py \
checkpoints/swinl_22k_w7_kitti.pth \
--show-dir swinl_22k_w7_kitti_results

📖 Citation

If you find this work useful in your research, please consider citing:

@misc{thelongrangedepth2025,
  title={Enhancing Long-Range Depth Estimation via Dual-Branch CNN-Transformer Integration and Multi-Dimensional Feature Fusion},
  author={Yunhao Chen and Qian Yin},
  journal={},
  year={2025},
  note={Submitted}
}

📩 Contact

For questions and issues, feel free to open an Issue or reach out via email: your_email@example.com