ZipMap: Linear-Time Stateful 3D Reconstruction via Test-Time Training

<h1 id="technical" Technical Analysis</h1 Jump to Section: Motivation motivation | Results summary-of-results | Conclusions final-conclusions "ZipMap" https://haian-jin.github.io/ZipMap/ is a stateful feed-forward model that achieves linear-time, bidirectional 3D reconstruction while matching or surpassing the accuracy of quadratic time methods. High-level Analogy : Imagine you're watching a very long documentary about a new city, and you want to create a perfect 3D model of that city. The Old Way Slow & Detailed : The best existing methods like VGGT or π3 are like having to re-watch the entire documentary from start to finish every single time you want to add a new building or detail to your 3D model. If the documentary gets longer, this re-watching process becomes incredibly, prohibitively slow – for a movie twice as long, it takes four times as much effort Another Old Way Fast & Less Detailed : Other methods like CUT3R or TTT3R try to be faster by only focusing on short segments of the documentary at a time, building up the city model piece by piece. But because they don't look at the whole picture efficiently, they might miss crucial connections between distant scenes, leading to a less accurate or complete 3D model. ZipMap's Way: ZipMap is like having a super-smart 'memory assistant' that watches the entire documentary once . As it watches, it continuously and very efficiently compresses all the essential information about the city into a single, compact, dynamic 'city g