Pavlo Puzikov
Menu

Dubai · 2026

Back to portfolio
Research Notes

Flip through the findings.

A field notebook of the research behind the work, rendered live in your browser. Each page is one finding. Turn the pages, or jump to any note below and read the full summary.

3D Gaussian Splatting

Radiance fields on a phone budget

Kerbl et al.'s 2023 method swaps NeRF's neural network for millions of explicit 3D Gaussians: it trains in minutes and renders above 60 fps. The work carries it from the paper to a web-native viewer that holds up inside a Safari-on-iOS budget, live in the BARNES Dubai property-search flow.

Method-paper quality, running on a mobile budget.

01 / 09Radiance fields on a phone budget

  1. Kerbl et al.'s 2023 method swaps NeRF's neural network for millions of explicit 3D Gaussians: it trains in minutes and renders above 60 fps. The work carries it from the paper to a web-native viewer that holds up inside a Safari-on-iOS budget, live in the BARNES Dubai property-search flow.

    Method-paper quality, running on a mobile budget.

    Splat Viewer

  2. Two dozen-plus 3DGS implementations each trade training speed against render fidelity differently. I integrated 20 engines across studio, hard, and city-scale scans, with the first 6 evaluations on PSNR, SSIM and LPIPS, plus a controlled raw-versus-cleaned comparison you can wipe through.

    Early evaluations already show the quality winner is not the speed winner.

    The 20-engine benchmark

  3. Splatting reconstructs surfaces from photographs, but a mirror or a floor-to-ceiling glass wall is a reflection, not a thing. Pipelines ghost the reflected room into the wrong volume or hollow out the geometry behind the glass. In luxury interiors, mirrors and glass are everywhere.

    Mirrors, glass and chrome are the dominant artefact class on hard scans.

    Research thread

  4. MirrorGaussian (Liu et al., 2024) models the real and mirror worlds at once, exploiting that a mirrored scene is symmetric about the mirror plane, with an intersection-aware formulation for partial reflections. Promising on benchmarks, not yet plug-and-play for production capture.

    The fix belongs at the photogrammetry layer, not the rasteriser.

    MirrorGaussian (arXiv 2405.11921)

  5. Rather than patch the renderer, the candidates sit upstream of reconstruction: mask-and-inpaint before COLMAP, dual-pass capture, and controlled-lighting subsets. The bet is that the capture problem has to be solved at the photogrammetry layer, where the bad data enters.

    Move the fix upstream of reconstruction.

    Research thread

  6. QLoRA, 4-bit NormalFloat with double quantization and paged optimisers, made adapting a 7-13B base on a single GPU tractable. Barnes Dubai LLM v8 is the production test of what a domain tune buys for one vertical: Dubai luxury real estate, mixed-language and terminology-heavy.

    Yes, on terminology and named-entity recall.

    Barnes Dubai LLM

  7. The instruction set, 15,369 pairs, is sourced from Dubai Land Department transactions, buildings, brokers, projects and developers, with UK, France and Singapore comparables for grounding. Tool-calling agents wrap the base with inventory, comp-search and valuation lookups.

    15,369 pairs across five public property registries.

    Barnes Dubai LLM

  8. The through-line across every thread: start at a canonical paper, run the benchmark or implementation step that actually closes the gap to production, and end at a live artifact that depends on the work. It makes the difference between using AI and shipping research visible.

    Three research lines, each with production credit.

    All research

  9. One handheld scan should locate safety assets as clickable 3D objects. The blocker is trajectory, not detection: the scanner exports only ~23s of poses out of six minutes, so the rest is re-derived from raw LiDAR and IMU. A KISS-ICP spike is metric on the stable segment but loses tracking full-scan.

    Detection was never the bottleneck. Recovering the full camera path is.

    Research thread

The threads, in full

From paper, through benchmark, to shipped.

Each thread starts at a canonical paper, runs through the benchmark or implementation step that closes the gap, and ends at a live artifact that depends on the work.

  1. Thread · 01

    3D Gaussian Splatting

    Real-time radiance fields applied to luxury real-estate visualisation.

    The Kerbl et al. 3DGS paper landed in 2023 and the field exploded into dozens of implementations, each making different trade-offs on training speed, render quality, and runtime fidelity. The benchmark closes the gap between method papers and what actually deploys to a Safari-iOS budget. The shipped artifact, Splat Viewer, is the answer running in production on the BARNES Dubai property-search flow.

    1. Paper

      3D Gaussian Splatting for Real-Time Radiance Field Rendering (Kerbl, Kopanas, Leimkühler, Drettakis, SIGGRAPH 2023)

      The seminal 3DGS paper. Replaces NeRF's MLP with millions of explicit Gaussians; trains in minutes, renders at 60+ fps.

      View abstract
      The authors introduce three elements that together achieve state-of-the-art visual quality while maintaining competitive training times and high-quality real-time (≥30 fps) novel-view synthesis at 1080p. Scenes are represented as 3D Gaussians initialised from sparse SfM points; an interleaved optimisation + adaptive-density-control step shapes the radiance field; a fast tile-based rasteriser keeps the render path GPU-resident. Demonstrated on several benchmarks against the prior NeRF state of the art.

    2. Benchmark

      20-engine 3DGS pipeline benchmark (6 evaluations so far)

      Around twenty Gaussian-splat engines wired into one harness across studio scans, hard scans, and city-scale capture, the first evaluations run. PSNR, SSIM, LPIPS, plus a controlled raw-vs-cleaned comparison the Operate-the-Result slider lets you wipe through.

    3. Shipped

      Splat Viewer in the BARNES property-search flow

      Web-native viewer with tap-to-move navigation, LOD streaming, and golden-hour relighting. First georeferenced building (Onyx Tower) live for brokers.

  2. Thread · 02

    Mirrors and glass capture

    The open problem Gaussian Splatting still hasn't solved for high-end interiors.

    Splatting reconstructs surfaces from photographs. Mirrors and floor-to-ceiling glass break that assumption, the surface is a reflection, not a thing, and current pipelines either ghost the reflected scene into the wrong volume or hollow the geometry behind the glass. For a luxury-real-estate use case (where mirrors and glass walls are everywhere), this is the bottleneck. The Living Homes industrial POV is that projector showrooms are the wrong proxy and that the problem has to be handled when the scene is captured, not corrected in the renderer afterward.

    1. Reference

      MirrorGaussian: reflective surface reconstruction (Liu et al. 2024)

      A leading attempt at handling mirrors in 3DGS by jointly modelling the real and reflected geometry. Promising on benchmarks; not yet plug-and-play for production capture.

      View abstract
      The key insight is that a mirrored scene is symmetric to the real scene with respect to the mirror plane. Building on this, the authors introduce a dual-rendering strategy that models the real-world and the mirror-world simultaneously, with an intersection-aware mirror Gaussian formulation that handles partial reflections. Demonstrated on real-world scenes containing mirrors with state-of-the-art reconstruction quality vs. prior 3DGS variants.

    2. In progress

      Capture-side mitigation pipeline

      Inside the benchmark, the per-engine failure analysis names mirrors / glass / chrome as the dominant artefact class on the Hard tier. Mitigation candidates are at the photogrammetry layer (mask + inpaint pre-COLMAP, dual-pass capture, controlled-lighting subsets) rather than the rasteriser.

  3. Thread · 03

    Domain LLM fine-tune

    Parameter-efficient adaptation of a base model to Dubai luxury real-estate.

    The QLoRA paper made it tractable to adapt a 7-13B base model on a single GPU. Barnes Dubai LLM v8 is the production answer to 'what does a domain-tuned model buy you versus base + RAG?' for a specific vertical (Dubai luxury real-estate, mixed-language, terminology-heavy, low-frequency entities). Trained on 15,369 instruction pairs sourced from Dubai Land Department records, UK Land Registry, France DVF, Singapore URA, and World Bank data.

    1. Paper

      QLoRA: Efficient Finetuning of Quantized LLMs (Dettmers, Pagnoni, Holtzman, Zettlemoyer, NeurIPS 2023)

      4-bit NormalFloat + double quantization + paged optimisers, the toolkit that brought 33B+ fine-tunes onto a single workstation GPU.

      View abstract
      QLoRA is an efficient finetuning approach that reduces memory usage enough to finetune a 65B-parameter model on a single 48GB GPU while preserving full 16-bit-finetune task performance. QLoRA backpropagates gradients through a frozen 4-bit quantized pretrained language model into Low-Rank Adapters (LoRA), and introduces three contributions: a new 4-bit NormalFloat data type, double quantization, and paged optimisers to manage memory spikes.

    2. Implementation

      15,369 instruction pairs, multi-source Dubai corpus

      Instruction set sourced from DLD transactions + buildings + brokers + projects + developers, plus UK/FR/SG comparables. Tool-calling agents (getCommunityInventory, comp-search, valuation lookups) wrap the base capability.

    3. Shipped

      Barnes Dubai LLM v8 in production

      FastAPI service the brokerage team dogfoods in a playground before WhatsApp distribution to brokers. The customer-facing answer to 'can a domain LLM beat base + RAG for our vertical' is currently 'yes, on terminology and named-entity recall'.

  4. Thread · 04

    Spatial asset intelligence

    Turning one handheld building scan into a per-object, queryable asset registry.

    A single walkthrough scan should do more than render a 3D model. It should locate every fire extinguisher, smoke detector and AC unit as a clickable object. The hard part is not detection, it is the camera trajectory: the device exports only ~23 seconds of poses out of a six-minute scan, so the rest is re-derived from raw LiDAR and IMU. A KISS-ICP spike is metric on the stable opening segment but loses tracking over the full walk, so IMU-fused LiDAR-inertial odometry is next.

    1. Reference

      KISS-ICP: In Defense of Point-to-Point ICP (Vizzo, Guadagnino, Mersch, Wiesmann, Behley, Stachniss, IEEE RA-L 2023)

      Point-to-point ICP against a sparse voxel map, with no IMU and no learned components. The dependency-light baseline for recovering a sensor trajectory from raw LiDAR sweeps.

      View abstract
      The system performs LiDAR odometry via point-to-point ICP combined with adaptive thresholding for correspondence matching, a robust kernel, motion compensation from a constant-velocity model, and point subsampling, working across sensors and mounting positions at sensor frame-rate without dataset-specific parameter tuning. The authors argue much of the complexity in modern odometry pipelines is unnecessary and a carefully engineered classic approach stays competitive.

    2. Reference

      FAST-LIO2: Fast Direct LiDAR-Inertial Odometry (Xu, Cai, He, Lin, Zhang, IEEE T-RO 2022)

      Tightly-coupled LiDAR + IMU via an iterated Kalman filter on an incremental kd-tree. The robustness step for the fast-rotation and featureless moments where LiDAR-only odometry drops.

      View abstract
      FAST-LIO2 is a direct LiDAR-inertial odometry framework that registers raw points to the map without feature extraction, tightly fusing inertial measurements through an iterated extended Kalman filter, and maintains the map with an incremental k-d tree (ikd-Tree) supporting incremental insertion and dynamic re-balancing. It demonstrates high accuracy and robustness at high odometry rates across structured and unstructured environments and varied LiDARs.

    3. In progress

      Trajectory recovery on the Onyx Tower reference scan

      On the reference scan a KISS-ICP spike recovers a metric path on the stable opening segment but loses tracking over the full six-minute walk; IMU-fused LiDAR-inertial odometry is the next step. The payoff is lifting per-frame object masks to per-object 3D across the whole scan, beyond the roughly 6 percent of assets currently anchored by direct camera evidence.