arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

Article | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Computer VisionEnglishPublished

AdaPixFlow learns a universal continuous‑time 3‑D motion model from sparse multi‑view RGB‑D data

November 21, 2026arXiv: 2611.26433v1

AdaPixFlow discovers a continuous‑time dynamics operator that works across different scenes, given only a few RGB‑D views. It fuses those views into a common voxel grid, uses a 3‑D U‑Net to predict a velocity field, and then integrates object points through a differentiable ODE solver to estimate motion between and beyond observation times.

Evaluated on held‑out PhysInOne scenes, AdaPixFlow increases the F1 score by 2.5 % over the best competing baseline for in‑between prediction and by 15 % for future prediction. Relative to per‑scene velocity‑field methods it outperforms them by 30 % and 85 % in the same tasks.

Table comparing AdaPixFlow against baseline methods in terms of F1 percentage for in-between and future predictions across several scenes, with the highest scores in bold and second-best underlined.
In-between predictionFuture prediction
Method
Input
Elastic Fall
DropIn Liquid
Crank Push
Panel Spin
Double Pendulum
Balloon Lift
Mean
Elastic Fall
Panel Spin
Double Pendulum
Balloon Lift
Mean
FreeGaveRGB, D0, M77.121.785.256.391.582.869.10.718.633.673.331.5
TRACERGB, D0, M0.015.170.232.515.745.829.92.532.834.444.028.4
SoMRGB-D, M, 1 view77.549.173.572.282.249.367.313.523.844.043.531.2
MoScaRGB-D, M, 1 view73.519.540.633.652.240.343.35.931.027.150.628.6
STv2RGB-D, M96.465.184.789.296.195.587.84.257.450.490.850.7
OpenD4RTRGB-D, M93.261.682.277.394.595.384.016.530.763.473.145.9
AdaPixFlowRGB-D, M97.377.987.286.395.495.890.022.047.680.783.258.4
Table 1. Quantitative F1 comparison on held‑out PhysInOne scenes, showing AdaPixFlow’s mean F1 of 90.0 (in‑between) and 58.4 (future) versus the next‑best STv2’s 87.8 and 50.7.Viswanath, Hrishikesh, Bera, Aniket · arXiv:2611.26433v1 · Table 1: · CC BY 4.0

The improvement is especially large for future steps that extend up to ten frames after the last input. Accuracy naturally degrades when observation spacing grows, and the model is less accurate on validation scenes featuring ground‑contact events compared to test scenes.

The performance gains are illustrated in Table 1, which reports per‑scene F1 scores for methods such as FreeGave, TRACE, and AdaPixFlow, and in Figure 3, which visualises the predicted 3‑D point clouds for in‑between and future prediction alongside ground truth and baseline methods.

Figure showing predicted 3D point clouds for different methods, with in-between predictions on the top row and future predictions on the bottom row, each panel labelled with the method name.
Figure 3. Visual comparison of predicted 3D point clouds for in-between and future steps, illustrating AdaPixFlow’s higher spatial accuracy relative to FreeGave and TRACE, alongside ground‑truth points.Viswanath, Hrishikesh, Bera, Aniket · arXiv:2611.26433v1 · Figure 3: · CC BY 4.0

Source paper

AdaPixFlow: Learning Continuous-Time 3D Dynamics Across Scenes from Sparse Multi-view Observations

Computer Vision
View on arXiv