# Gallileo-4D: Frozen Backbone Ensemble for Dynamic 4D Reconstruction

arXiv: https://arxiv.org/abs/2608.19743 · HTML: https://mlsanity.com/paper/2608.19743/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 29 of its 34 references that are in the map.
- First posted on arXiv 20 Aug 2026.

## Builds on (most cited first, 29 of 29)

- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [DINOv2: Learning Robust Visual Features without Supervision](https://mlsanity.com/paper/2304.07193/) (2023)
- [3D Gaussian Splatting for Real-Time Radiance Field Rendering](https://mlsanity.com/paper/2308.04079/) (2023)
- [Overcoming catastrophic forgetting in neural networks](https://mlsanity.com/paper/1612.00796/) (2016)
- [Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles](https://mlsanity.com/paper/1612.01474/) (2016)
- [DUSt3R: Geometric 3D Vision Made Easy](https://mlsanity.com/paper/2312.14132/) (2023)
- [VGGT: Visual Geometry Grounded Transformer](https://mlsanity.com/paper/2503.11651/) (2025)
- [Do ImageNet Classifiers Generalize to ImageNet?](https://mlsanity.com/paper/1902.10811/) (2019)
- [Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time](https://mlsanity.com/paper/2203.05482/) (2022)
- [Grounding Image Matching in 3D with MASt3R](https://mlsanity.com/paper/2406.09756/) (2024)
- [PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning](https://mlsanity.com/paper/1711.05769/) (2017)
- [Robust fine-tuning of zero-shot models](https://mlsanity.com/paper/2109.01903/) (2021)
- [Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution](https://mlsanity.com/paper/2202.10054/) (2022)
- [Depth Anything 3: Recovering the Visual Space from Any Views](https://mlsanity.com/paper/2511.10647/) (2025)
- [CoTracker: It is Better to Track Together](https://mlsanity.com/paper/2307.07635/) (2023)
- [MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motion](https://mlsanity.com/paper/2410.03825/) (2024)
- [Kubric: A scalable dataset generator](https://mlsanity.com/paper/2203.03570/) (2022)
- [π3: Permutation-Equivariant Visual Geometry Learning](https://mlsanity.com/paper/2507.13347/) (2025)
- [PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point Tracking](https://mlsanity.com/paper/2307.15055/) (2023)
- [Shape of Motion: 4D Reconstruction from a Single Video](https://mlsanity.com/paper/2407.13764/) (2024)
- [SpatialTracker: Tracking Any 2D Pixels in 3D Space](https://mlsanity.com/paper/2404.04319/) (2024)
- [Better Aggregation in Test-Time Augmentation](https://mlsanity.com/paper/2011.11156/) (2020)
- [St4RTrack: Simultaneous 4D Reconstruction and Tracking in the World](https://mlsanity.com/paper/2504.13152/) (2025)
- [TAPVid-3D: A Benchmark for Tracking Any Point in 3D](https://mlsanity.com/paper/2407.05921/) (2024)
- [Any4D: Unified Feed-Forward Metric 4D Reconstruction](https://mlsanity.com/paper/2512.10935/) (2025)
- [V-DPM: 4D Video Reconstruction with Dynamic Point Maps](https://mlsanity.com/paper/2601.09499/) (2026)
- [Trace Anything: Representing Any Video in 4D via Trajectory Fields](https://mlsanity.com/paper/2510.13802/) (2025)
- [4RC: 4D Reconstruction via Conditional Querying Anytime and Anywhere](https://mlsanity.com/paper/2602.10094/) (2026)
- [Syn4D: A Multiview Synthetic 4D Dataset](https://mlsanity.com/paper/2605.05207/) (2026)

## Abstract (arXiv)

We describe our entry to the PhysAI Dynamic 4D Reconstruction Challenge, which placed third of 27 teams at 0.58356 APD on the final leaderboard, without a single gradient update. This was not the plan: of thirteen fine-tuning configurations of a pre-trained 4D backbone, twelve degraded the challenge score, and eleven of those twelve improved local validation at the same time. We trace this inversion to the structure of the benchmark: only 25% of the evaluation set belongs to the data variant released for training, so updates that fit the available data damage the pre-trained features the remaining 75% relies on. Our system therefore freezes the backbone and spends its budget at inference time, fusing three decoding configurations -- temporal stride-3, horizontal-flip test-time augmentation, and dense stride-1 -- under a convex weighting. The ensemble recovers +0.041 APD over the frozen baseline, more than any training run achieved, at zero training cost.