3D Object Reconstruction with TensoRF

---

This project implements efficient 3D reconstruction and novel view synthesis using TensoRF, a tensor decomposition-based variant of Neural Radiance Fields (NeRF).
TensoRF significantly accelerates training and rendering by replacing NeRF’s dense MLP with compact tensor factorization, while maintaining high-quality reconstruction.
We use this framework to reconstruct real-world objects from multi-view images captured with a handheld smartphone, and visualize the reconstruction via novel view rendering.

📦 Dataset Access

The real-world dataset used in this project, along with COLMAP-processed poses and the resulting outputs, can be downloaded via the following link:

🔗 Baidu Netdisk Download – TensoRF Dataset and Results (extraction code: 9buk)

🔗 Official TensoRF repo: https://github.com/apchenstu/TensoRF

---

📁 Project Structure

🔧 Main Components

• train.py
  The main training script that initializes the tensor field, sets up loss functions, and performs optimization over iterations.

• renderer/
  Contains rendering utilities and volumetric rendering logic using the decomposed tensor components.

• models/
  Implements the TensorVMSplit and TensorCP models used for representing the radiance field via low-rank tensor factorization.

• opt/
  Contains optimization tools, learning rate schedulers, and loss functions, including:

  • TV Loss for regularizing appearance and density fields.
  • SSIM and PSNR tracking for quality evaluation.

• configs/
  YAML configuration files for different experiments (resolution, backbone, dataset, loss weights, etc.).

• scripts/
  Shell scripts or command-line examples for launching training on different datasets.

• test/
  Contains rendering and evaluation scripts for novel view synthesis after training.

---

🏗️ Workflow

📸 (1) Data Preparation

• Capture multi-view images of an object with overlapping views using a handheld smartphone.
• Use COLMAP to compute accurate camera poses.
• Organize the images and camera intrinsics in NeRF-compatible format (e.g., LLFF-style or synthetic layout).
• Resize images to a consistent resolution (optional, for speed).
• Split dataset into training and test sets (e.g., 80/20).

🧠 (2) Model Training

• Edit the YAML config (e.g., configs/hotcrush_tensorf.yaml) to specify:

  • Dataset path
  • Model type: TensorVMSplit or TensorCP
  • Number of iterations
  • Loss weights for TV, appearance, and density
  • Logging intervals

• Run training:

python train.py --config configs/hotcrush_tensorf.yaml