DRoPS: Dynamic 3D Reconstruction of Pre-scanned Objects

Evaluation on the CMU Panoptic Studio dataset [4], which features real-world multi-view captures of complex human motions such as juggling, softball, and tennis. We use a single camera view for training and evaluate novel view synthesis on 30 held-out cameras at extreme viewing angles. We compare against state-of-the-art optimization-based (HiMoR [1], OriGS [2]) and generative (Cog-NVS [3]) approaches. DRoPS significantly outperforms the baselines in maintaining a consistent object geometry and sharp appearance, while accurately modeling the scene dynamics.

See more examples here.

Evaluation on our synthetic dynamic animals dataset generated using TrueBones [5] animated characters. Since the data is synthetic, we have access to ground-truth depth and camera parameters, allowing us to validate our approach under controlled conditions. We compare against state-of-the-art optimization-based (HiMoR [1], OriGS [2]) and generative (Cog-NVS [3]) approaches. DRoPS significantly outperforms the baselines in maintaining a consistent object geometry and sharp appearance, while accurately modeling the scene dynamics.

See more examples here.

DRoPS can operate on casual, fully monocular videos where no multi-view pre-scan is available. Instead, the pre-scan is generated from the first video frame using an off-the-shelf image-to-3D model. As shown below, DRoPS achieves high-quality, complete dynamic 3D reconstruction even in this challenging in-the-wild setting. Videos are taken from the DAVIS dataset [6] and the internet.

We ablate the key design choices of our method on the CMU Panoptic Studio dataset [4]. Each row shows the result of removing one component on a held-out test camera. Removing any key component significantly degrades reconstruction quality, highlighting their contribution.

See the full ablation study here.

DRoPS produces an explicit dynamic 3D representation from which long-range 3D point trajectories emerge naturally. We compare our 3D tracking against optimization-based baselines HiMoR [1] and OriGS [2], which also incorporate explcit dynamic 3D representations with deforming Gaussian splats. The ground-truth positions are visualized as crosses, the predictions as circles, and the lines connecting to them show the error.
DRoPS achieves significantly more accurate long-range 3D tracking than the baselines, while maintaining a consistent object geometry and sharp appearance. See our paper for quantitative evaluation and more details.

DRoPS enables Text-to-4D generation: given a text prompt describing a dynamic subject, we first generate a video using an off-the-shelf text-to-video model, then apply DRoPS with the in-the-wild pipeline for dynamic 3D reconstruction. The colored lines visualize 3D point trajectories emerging from our representation.

@inproceedings{drops2026,
    author = {Narek Tumanyan, Samuel Rota Bulò, Denis Rozumny, Lorenzo Porzi, Adam Harley, Tali Dekel, Peter Kontschieder, Jonathon Luiten},
    title = {DRoPS: Dynamic 3D Reconstruction of Pre-scanned Objects},
    year = {2026},
}