quang/ai
VI Hire me
Top 4 · 1 of 12 finalists 3D VisionTSC · Team Zillexa

BTS Digital Twin — Viettel AI Race 2026

Novel view synthesis of telecom (BTS) towers from drone captures with 3D Gaussian Splatting — then large urban scenes under an inference-time limit on an NVIDIA H200 in the final round.

of 1,600+ teams
Top 4
finalist teams
12
training GPUs
4×A5000
final-round GPU
H200

The challenge

Viettel AI Race 2026, Track 1 asked teams to build a digital twin of BTS telecom towers: reconstruct each site from drone imagery and render views the cameras never captured. The final round moved to large urban scenes and added an inference-time limit, so quality had to be traded against render time on an NVIDIA H200.

Approach

  • gsplat MCMC backbone for 3D Gaussian Splatting.
  • Camera-model plumbing: SIMPLE_RADIAL cameras handled by undistorting, training in pinhole, then re-distorting renders.
  • Ensembles: multiple backbones fused with SELECT-median.
  • Training infrastructure: gsplat MCMC training sharded across 4×A5000 with a merge step and a monitor for idle GPUs.

Final round: large scenes under a time budget

  • Capped MCMC, weight decay and MVGS-style multi-view gradient accumulation.
  • A fine-tuned SCUNet restoration pass.
  • Count-based pruning to trade quality against render time.
  • A Hopper (sm_90) Docker build for the H200.

Debugging the rasterizer

At full resolution, training broke inside gsplat's CUDA projection. An audit traced it to an int32 intersection overflow; it was fixed by clamping the projected radius.

Result

Team Zillexa finished Top 4 of 1,600+ teams (Prospective Prize) and was one of 12 finalist teams. The code is private; I'm happy to walk through the pipeline in an interview.

→ More work

Other case studies

iPBMS — real-time patient monitoring

Fall and seizure detection on live RTSP hospital camera streams — from pose model to on-premise production on two cameras, with alerts to care staff in under two seconds.

 Inference-Efficient ConvMAE

Can a masked-autoencoder backbone be cheaper at inference without losing accuracy? Four hybrid designs, one fair protocol, measured under TensorRT rather than guessed from parameter counts.

 Face-Aware MAE

Identity-aware masked autoencoding: instead of masking random blocks, point MAE pre-training at the eyes, nose, mouth and central facial structure — and change nothing else.

 Vietnamese License Plate Recognition Benchmark

A reproducible benchmark for Vietnamese automatic license plate recognition: three YOLO generations, three OCR pipelines, real RTSP traffic-camera data.
esc