acceptodds
Withdrawn from ICLR 2027

EgoLite-G: Restoring the Physical Gauge in Egocentric Token World Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Token-based egocentric world models canonicalize every modality in its own first-frame camera frame. We show that this design silently removes the physical gauge—the gravity direction and the height of the wearer above the support surface—on which world-frame body pose depends. We introduce EgoLite-G, a gauge-state egocentric world model that keeps the gauge as an explicit, persistent state: gravity from the head-worn IMU and a take-level support height estimated from the head trajectory alone. The state enters a masked-token world model through zero-initialized gauge embeddings and is trained with privileged, noise-perturbed heights so that the same network is deployed with legal estimates. On the official EE4D-Motion validation set (99 unseen participants, 44,046 windows, scored once after all predictions were frozen), EgoLite-G lowers world MPJPE to 130.4 mm: 17.2 mm below a budget-matched fine-tuned ReViV, 19.3 mm below the strongest pre-registered baseline, and 5.6 mm below EgoAllo given the very same floor estimate, while matching EgoAllo supplied with a label-derived floor. Controlled studies show that the gain is caused by the gauge itself: shuffling gravity at inference costs 15.9 mm, and body error decreases monotonically as the gauge becomes more accurate.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Egocentric Vision, 3D Human Pose Estimation, Whole-Body Reconstruction, Geometric Calibration, Geometric Decoding

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_nFvUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies full-body motion reconstruction in world coordinates from head-mounted IMU measurements. It estimates the gravity direction and the floor height, then injects these two physical properties into a pretrained/frozen ReViV.

Critical strengths
  1. It gives many implementation details.
Critical weaknesses
  1. The paper title is misleading: This paper is essentially full-body mesh reconstruction under the world coordinate. It is not a world model that can predict future states or model action-conditioned dynamics.

  2. The novelty is unclear: The proposed gauge mainly includes gravity and floor estimation. However, the gravity is already available in the original world frame and does not need to be predicted. Moreover, floor height estimation has been studied in many prior works, such as EgoAllo.

  3. Many equations are not defined, e.g., z_{1}^{zpf}, making the methodology difficult to read.

  4. Unclear expression in the experimental setup. For example, in 4.1 Baseline, the expression “ReViV fine-tuned with exactly our budget.” fails to present any information details.

Questions for authors

No

Reviewer_HSmWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EgoLite-G, an egocentric world model that explicitly keeps gauge state. Prior egocentric world model approaches neglected the physical gauge (gravity direction and floor height information), creating pose ambiguity. EgoLite-G keeps gauge information estimated from an IMU device in an explicit state. Further, it is trained with privileged, noise-perturbed heights to imitate noisy practical conditions. The proposed method is evaluated on the EE4D-Motion dataset and reduces world MPJPE compared to ReViV (the baseline model) and EgoAllo. Experiments further strengthen the claim that the gain is from gauge information.

Critical strengths
  • The paper points out a critical issue of gauge information loss in the existing method and proposes a method to deal with it.
  • Thorough evaluation with various statistics, including multiple seeds and confidence intervals provided for comparison.
Critical weaknesses
  • The contribution of uncertainty-aware training is not clear; in experiments, clean training rather gives better results than noisy or signal-aware training.
  • Gauge state is given for each take; this would limit application in case ground height is variable.
  • Lack of qualitative comparison makes it hard to evaluate the comparative advantage of this method clearly.
Questions for authors
  • The role of the noisy training should be clarified to attribute it as a contribution; is there any evidence to support it?
  • Can authors provide a qualitative comparison in the evaluation section to prove its practical advantage?
Reviewer_99vqOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript proposes EgoLite-G for 3D human motion reconstruction from egocentric head trajectories using the pretrained ReViV model. The authors argue that first-frame camera canonicalization removes the gravity direction and floor-relative head height, and propose to restore these quantities as an explicit gauge state. Gravity is obtained from the IMU-aligned camera pose, while floor height is estimated from head-trajectory statistics.

Critical strengths

None

Critical weaknesses

W1. Inconsistencies with prior work: The manuscript makes several claims that I could not verify. First, the introduction refers to a previous "EgoLite line of work" (L50) without providing any reference. Second, the central motivation claims that ReViV (ECCV 2026) removes gravity information through first-frame canonicalization, whereas ReViV constructs its body reference frame from the first camera pose with the up-vector aligned opposite to gravity. It is therefore unclear why gravity information would be lost in the proposed setup. These discrepancies raise concerns about the correctness of the technical description and the central motivation of the paper.

W2. Reliability of the experimental evaluation: The paper reports results on an "official EE4D-Motion validation set" with 99 participants and 44,046 windows, whereas UniEgoMotion (ICCV 2025) reports approximately 4,400 validation samples. The relationship between these protocols is unclear. Additionally, the baseline results differ between Tabs. 4 and 7 (e.g., ReViV: 149.75 vs. 156.45 mm) without a clear explanation. Surprisingly, the manuscript also does not provide a single qualitative visualization of the reconstructed 3D human motion. These issues raise concerns about the reliability of the experiments.

Questions for authors

Q1. Can the authors clarify the discrepancies with the original ReViV representation and implementation (see W1), and provide a reference for the claimed previous EgoLite work?

Q2. How does the reported validation split relate to the official EE4D-Motion protocol? Additionally, can the authors explain the differences in baseline results between Tabs. 4 and 7 and provide qualitative comparisons of the reconstructed 3D human motion?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.