Akshay Kambli

Computer Vision · 3D Reconstruction

Recon3D

Vision-based 3D terrain mapping that replaces LiDAR with a camera and a depth model — video in, dense point cloud out, at a fraction of survey cost.

Recon3D

Organisation

Independent Project

Role

Solo — pipeline, reconstruction, web app

Timeline

2026

Stack

Python PyTorch Depth Anything 3 Open3D OpenCV NumPy Flask
View repository

Overview

Accurate 3D terrain mapping matters for mining, construction and infrastructure planning, and today it runs on LiDAR. Mapping roughly 10 sq km can cost ₹50–60 lakhs, which puts frequent surveys out of reach for most organisations and rules out regular monitoring entirely.

Recon3D goes after that cost by replacing the sensor with software. A camera on a drone, UAV or ground vehicle captures video; the pipeline turns it into a dense 3D point cloud using monocular depth estimation. The hardware gets cheap and the intelligence moves into the model.

Key Features

  • Video-to-point-cloud pipeline built on Depth Anything 3 (DA3NESTED-GIANT-LARGE-1.1), converting monocular RGB footage into metric-like depth and then into structured 3D geometry, with an optional higher-accuracy ray-pose mode.
  • Quality-aware frame extraction — frames are sampled at a target FPS and scored for sharpness by Laplacian variance, so blurry footage is filtered out before it can poison the reconstruction.
  • Memory-bounded processing with configurable frame caps and resize width, tuned so full reconstructions run on a single T4-class GPU rather than requiring a workstation.
  • Point cloud generation and export through Open3D with voxel downsampling, writing both PLY and GLB so output drops straight into standard 3D and web viewers.
  • Flask web application for uploading footage and inspecting results, alongside a local runner for batch processing off a workstation.

Impact

A LiDAR survey is a budget decision; a camera pass is not. By moving the hard part into depth estimation, Recon3D targets terrain reconstruction that can be repeated as often as the site changes, including in remote regions where hauling survey hardware is the real constraint.

Status: the depth-to-point-cloud pipeline is built and integrated. The custom long-endurance drone platform, self-tuning PID control, GPS-denied visual-inertial navigation, and the “Chain of Zoom” super-resolution stage for denser distant geometry are all in development — they are the roadmap, not shipped capability.

← Back to work