Get3D Knowledge Center

03 Vehicle Removal

3 min read

Key Functions: Moving Vehicle Removal

Summary: Moving vehicles create ghost artifacts in photogrammetric 3D models that can corrupt 20-30% of road surfaces in dense urban areas. Get3D Mapper solves this through a two-stage AI pipeline — detection and multi-view consistency verification — that removes moving vehicles and inpaints clean road surfaces automatically, with no manual correction required.

Who This Guide Is For: Photogrammetry operators and production managers dealing with traffic-heavy urban capture sites where ghost vehicle artifacts are unavoidable.

One of the most persistent quality problems in large-scale urban 3D reconstruction is the presence of moving vehicles. Cars, trucks, and buses captured at different positions across overlapping images produce a well-known class of artifacts: ghosting, duplicated geometry, and discontinuous road surfaces.

Content-aware reconstruction eliminates these artifacts at the source — during the reconstruction pipeline itself, not as a manual clean-up step afterward.

The Problem: Ghost Vehicles in Multi-View Reconstruction

Standard photogrammetric pipelines assume the scene is static. Every point in the 3D model is computed by triangulating the same physical feature across multiple images. When a vehicle moves between exposures, it appears at different positions in each image — and the algorithm cannot correctly triangulate it.

The result is characteristic ghost geometry: smeared, semi-transparent-looking vehicle shapes fused into the road surface, duplicated car silhouettes, and disrupted texture continuity across the road.

In dense urban environments where traffic is continuous, this can corrupt a significant portion of the road mesh — making it unusable for pavement inspection, lane detection, or any application that requires an accurate ground surface.

How Content-Aware Removal Works

Get3D Mapper addresses this through a two-stage pipeline combining AI-based object detection and multi-view geometric consistency.

Stage 1 — AI Detection

A deep learning detection model identifies vehicle instances across all input images before reconstruction begins. Rather than working on the 3D output, the detection operates directly on the 2D image space, where vehicle identification is both faster and more reliable.

Each detected vehicle is masked out of the images it appears in.

Before: Vehicle removal - before

After: Vehicle removal - after

Stage 2 — Multi-View Consistency Verification

Detection alone is not sufficient — false positives, partial occlusions, and lighting variation can cause misses or over-masking. A multi-view consistency check cross-validates detections across overlapping views, ensuring that:

  • A region masked as a vehicle in one image is confirmed by geometry inconsistency across neighboring images
  • Static objects (e.g., parked vehicles in exactly the same position across all images) are preserved rather than incorrectly removed
  • The background road surface is inpainted from adjacent clean views where the vehicle was absent

Vehicle removal multi-view consistency

Result

The output is a clean road mesh with no ghost geometry, restored surface continuity, and accurate texture — ready for downstream analysis without any manual correction.

Source: Get3D Knowledge Center

Graph View