ANTARA: 3D Reconstruction for Single-Pass Drone Video

ANTARA is a high-speed, edge-deployable AI pipeline that converts single-pass monocular drone videos into metric-accurate 3D point clouds, solid meshes (OBJ), and Georeferenced Orthomosaics (GeoTIFF) without the need for Ground Control Points (GCPs).

Built for the Smart India Hackathon 2025, this pipeline targets sub-meter accuracy and processes 10-minute 4K video feeds in under 15 minutes using true multiprocessing and Vision Transformers.

---

Step-by-Step Installation Guide (Windows / Linux)

Because this pipeline integrates complex C++ compilations (3D Gaussian Splatting) and heavy AI models, you must follow these steps precisely.

Phase 1: System-Level Software (Install these FIRST)

You cannot just use pip install for this project. Your operating system needs the correct compilers and rendering engines installed first.

1. NVIDIA CUDA Toolkit (Required for 3DGS & PyTorch)

• Official Link: CUDA Toolkit 11.8 Archive
• What to do: Download and install the version for your OS.
• Crucial ANTARA Warning: The official NVIDIA site will try to push CUDA 12.x on you. Do NOT install CUDA 12.x. The 3D Gaussian Splatting rasterizer is highly unstable on Windows with CUDA 12. You must install version 11.8 to compile the C++ extensions error-free.

2. Visual Studio 2022 C++ Build Tools (Windows Only)

• Official Link: Visual Studio Build Tools
• What to do: Install it and select the "Desktop development with C++" workload.
• Crucial ANTARA Warning: The nvcc CUDA compiler needs this to build the SuGaR and 3DGS meshes.

3. COLMAP (Required for Camera Initialization)

• Official Link: COLMAP GitHub Releases
• What to do:
  • Windows: Download COLMAP-x.x-windows-cuda.zip. Extract it to C:\COLMAP. You must add this folder to your Windows System PATH environment variable.
  • Linux: Run sudo apt-get install colmap
• Crucial ANTARA Warning: Our pipeline uses SLAM3R for the point cloud, but the 3DGS engine strictly requires COLMAP's cameras.bin files to start training.

4. FFmpeg (Required for Video Extraction)

• Official Link: FFmpeg Download (For Windows, use Gyan.dev)
• What to do: Extract the ZIP and add the bin/ folder to your System PATH. Verify by typing ffmpeg -version in your terminal.

5. Miniconda (Required Python Manager)

• Official Link: Miniconda3 Installers
• What to do: Install the 64-bit version. We strongly advise against using standard Python/pip, as it routinely corrupts PyTorch CUDA environments.

---

Phase 2: Project Setup

Once the system prerequisites are installed, open your Anaconda Prompt (Windows) or Terminal (Linux) and run:

1. Clone the Repository Because this project links to external AI repositories (SAM2, SLAM3R, 3DGS), you must use the --recursive flag:

git clone --recursive https://github.com/Yoy20-New/ANTARA-3D-Reconstruction-for-Single-Pass-Drone-Video.git
cd ANTARA-3D-Reconstruction-for-Single-Pass-Drone-Video

2. Create the Conda Environment

conda create -n antara python=3.10 -y
conda activate antara

3. Install PyTorch 2.x (CUDA 11.8) Do not install standard PyTorch. Force the CUDA 11.8 version:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

4. Install Project Requirements

pip install -r backend/requirements.txt

5. Download AI Weights & Apply Patches We have written automated scripts to fetch the heavy AI models (like YOLO and SAM2) and apply our custom TensorRT modifications to the submodules.

python scripts/download_weights.py
python scripts/apply_patches.py

---

Running the Pipeline

You can run the full pipeline in two ways:

1. Headless CLI Mode:

python run_pipeline.py --video videos/drone_footage.mp4 --gps videos/DJI_gps.SRT --output output/

2. Interactive Web UI: Start the FastAPI background server:

python backend/server.py

Then navigate to http://localhost:8000 to upload videos and view live metrics in the 3D Geographic HUD.

---

Built for SIH 2025. This project utilizes code from SLAM3R, Meta Segment Anything 2, and Inria Gaussian Splatting.