TRIG: Trajectory-Rig Decoupled Metric Geometry Learning

Chang Huang; Handong Wang; Lirong Yang; Lizhou Liao; Shuai Yang; Weiwei Liu; Wentao Xu

Splitting camera motion from rig shape yields metric 3D from cameras alone

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-08 23:35 UTC pith:TMPJLIEF

load-bearing objection TRIG: solid architectural contribution for multi-camera driving geometry, but internal table inconsistency undermines the ablation evidence the 2 major comments →

arxiv 2607.05801 v2 pith:TMPJLIEF submitted 2026-07-07 cs.CV cs.RO

TRIG: Trajectory-Rig Decoupled Metric Geometry Learning

Lizhou Liao , Wentao Xu , Handong Wang , Lirong Yang , Shuai Yang , Weiwei Liu , Chang Huang This is my paper

classification cs.CV cs.RO

keywords geometrydrivingmetrictrigautonomousdecoupledego-motionestimation

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

TRIG argues that a multi-camera driving system's pose can be cleanly factored into two parts: a time-varying ego-motion trajectory and a fixed camera-rig topology. Existing visual geometry models entangle these, losing the ability to exploit vehicle-side geometric priors and requiring post-hoc scale alignment. By encoding and supervising trajectory and rig separately, and by routing them through a sparse attention scheme that handles cross-camera spatial reasoning only at selected layers while continuously aggregating temporal cues, TRIG claims to predict metric-scale depth, 3D structure, and ego-pose directly from images without any Sim(3) alignment step. The paper reports substantial improvements over prior methods across five autonomous driving benchmarks.

Core claim

The central claim is that decoupling a rigid multi-camera system's pose into ego-trajectory and camera-rig components, through separate encoding, sparse attention, and split supervision, eliminates metric-scale ambiguity and produces directly metric geometry from images alone. The trajectory-rig factorization (Eq. 2) is the load-bearing identity: each camera pose is the product of a per-timestep ego-vehicle pose and a fixed per-camera extrinsic. This decomposition lets the model treat temporal motion and static spatial layout as independent information streams, which the paper shows is sufficient to define a deterministic metric scale reference without external alignment.

What carries the argument

trajectory-rig pose factorization

Load-bearing premise

All baseline numbers are taken from the DVGT paper rather than re-implemented under TRIG's own training and evaluation pipeline, so the very large reported improvements could be inflated if there are differences in input resolution, preprocessing, or metric computation between the two setups.

What would settle it

Re-running the strongest baseline (DVGT) under TRIG's exact evaluation protocol, input resolution, and preprocessing. If the gap shrinks substantially, the core claim of large improvement from decoupling weakens.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

If the factorization generalizes, any rigid multi-camera platform (drones, robots, underwater rigs) could adopt the same trajectory-rig split to get metric geometry without LiDAR or post-hoc alignment.
The sparse attention pattern—rig blocks at 5 of 24 layers, trajectory blocks everywhere—suggests that cross-camera spatial reasoning is needed far less frequently than temporal aggregation, which could inform efficient transformer design for other multi-sensor fusion tasks.
Direct metric prediction without Sim(3) alignment would simplify the perception-to-planning pipeline, removing a runtime post-processing step that depends on sparse depth supervision.
The decoupled supervision strategy (separate rig and trajectory relative-pose losses) could be applied to other pose-conditioned geometry models as a drop-in training improvement.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

TRIG: solid architectural contribution for multi-camera driving geometry, but internal table inconsistency undermines the ablation evidence

read the letter

TRIG factorizes multi-camera poses into ego-trajectory and static rig components, builds a sparse attention scheme around that factorization, and reports large improvements over prior work on five driving benchmarks. The core idea — that you should not entangle time-varying ego-motion with fixed camera extrinsics when conditioning a geometry model — is sound and well-motivated for the autonomous driving setting. The architectural instantiation is clean: separate trajectory and rig embeddings composed additively, rig blocks sparsely inserted at 5 of 24 layers, and a decoupled supervision scheme that splits relative pose loss into cross-camera and cross-time terms. The ablations in Tables IV, VI, and VII are internally consistent with each other and show that each component contributes. The runtime comparison (Table V) showing 1.4–1.7x speedup over VGGT-style attention is a nice practical point. Credit is due for a well-structured engineering contribution that addresses a real gap in applying visual geometry foundation models to driving. The trajectory-rig decomposition itself is standard SE(3) kinematics, not a new insight — the contribution is in how it is operationalized within the architecture and training objective. Now the problems. The most serious issue is the discrepancy between the ablation tables and the main results tables. Table VII's 'Full model' numbers do not match Tables I–III's 'TRIG (Ours)' for the same datasets and metrics. For example, KITTI Acc is 0.291 in Table VII but 0.333 in Table I; Waymo Comp is 0.879 vs 1.192; DDAD AbsRel is 0.099 vs 0.109. The ablation numbers are consistently better than the main results. The paper gives no explanation — no mention of different checkpoints, evaluation settings, or frame counts. This matters because the ablations are the evidence that each component earns its place. If the 'full model' in the ablations is not the same model reported in the main tables, the causal link between the architectural contributions and the SOTA claims is weakened. The second concern is baseline fairness: all baselines are taken from the DVGT paper rather than re-run under identical conditions. Given the very large reported margins (e.g., KITTI Acc dropping from 0.846 to 0.333), a controlled comparison would substantially strengthen the claims. No code or weights are released. These are the kind of issues that a serious referee should press on, but they are fixable — the paper does not appear to be fundamentally broken. The architectural design is reasonable, the ablation structure is right (just needs consistent numbers), and if even half the reported improvements hold under controlled comparison, this is a useful contribution for researchers working on multi-camera driving perception. Recommend sending to peer review. The table inconsistency and baseline fairness need to be resolved before acceptance, but the work deserves a careful read.

Referee Report

2 major / 7 minor

Summary. The paper proposes TRIG, a framework for multi-camera autonomous driving geometry perception that factorizes camera poses into ego-trajectory and camera-rig components (Eq. 2), encodes them via separate MLPs with additive composition (Eq. 5), and introduces sparse Temporal-Spatial Attention (STSA) that interleaves Rig Blocks and Traj Blocks (Eq. 7). Decoupled pose supervision (Eqs. 8-9) separately constrains cross-camera rig geometry and cross-time ego-motion. Built on VGGT with a frozen DINOv2 backbone, TRIG is evaluated on five benchmarks (KITTI, NuScenes, Waymo, OpenScene, DDAD) across 3D reconstruction, metric depth, and ego-pose estimation, reporting substantial improvements over prior methods including DVGT.

Significance. The trajectory-rig factorization is a clean and well-motivated decomposition for the autonomous driving setting, and the sparse attention design that separates cross-camera from temporal interactions is a sensible efficiency measure (Table V reports 1.43x-1.72x speedup). The paper provides ablation studies isolating each architectural component (Tables IV, VI, VII) and reports direct metric-scale prediction without post-hoc Sim(3) alignment, which is a practically useful property. The scope of evaluation across five benchmarks and three tasks is commendable. However, the significance of the reported gains is difficult to assess due to internal inconsistencies in the results tables and the absence of self-run baselines, as detailed below.

major comments (2)

Tables IV, VI, and VII report 'Full model (Ours)' values that systematically differ from the 'TRIG (Ours)' entries in the main results Tables I-III for the same datasets and metrics. Concrete examples: KITTI Acc is 0.333 in Table I vs 0.291 in Table VII (13% discrepancy); Waymo Comp is 1.192 in Table I vs 0.879 in Table VII (26% discrepancy); DDAD AbsRel is 0.109 in Table II vs 0.099 in Table VII (9% discrepancy); KITTI AbsRel is 0.046 in Table II vs 0.041 in Table VII (11% discrepancy). The ablation tables consistently show better numbers than the main results. If the 'full model' in ablations is not the same checkpoint or evaluation protocol as 'TRIG (Ours)' in the main tables, then the ablation deltas (e.g., removing Rig blocks increases Acc from 0.411 to 0.881 in Table IV) may not reflect the actual contribution of each component to the reported SOTA numbers. The paper provides no解释.
§V, first paragraph: the statement that 'baseline results are taken from the DVGT paper' means TRIG's numbers are compared against published values from another paper rather than re-implemented baselines under identical training and evaluation conditions. This is load-bearing because the reported improvements are very large (e.g., KITTI Acc improving from 0.846 to 0.333, a ~2.5x reduction; KITTI AUC@30° from 87.6 to 99.1). If there are differences in evaluation protocol, input resolution, data preprocessing, or metric computation between the DVGT paper's setup and TRIG's setup, the magnitude of improvement could be inflated. The paper should either re-run at least the strongest baseline (DVGT) under its own pipeline or explicitly confirm that all evaluation settings are identical and document what those settings are.

minor comments (7)

Eq. (8): the notation uses both T^{(k),t}_{c_i←c_j} and T^{t,c_i}_{w→c} styles. Please unify the notation for predicted vs ground-truth transforms and the direction of subscripts throughout.
§IV-B: the choice of reference camera and reference timestamp for extracting trajectory and rig components is mentioned but not justified. A brief note on why the front camera at t=0 is a reasonable reference would strengthen the section.
Table I: the 'Metric' column marks some methods as requiring post-hoc Sim(3) alignment. It would help to briefly state in the caption or table notes what alignment method each non-metric method uses (Umeyama with sparse LiDAR? with monodepth?).
Fig. 2 caption: 'The original VGGT DPT and camera heads produce the final depth and pose predictions, and are omitted for clarity.' This is fine, but the figure should still indicate where the depth and pose outputs come from, even schematically, as the current figure is otherwise incomplete.
§IV-C, Eq. (7): the Rig Block layer indices {0, 4, 11, 17, 23} are stated without justification. A brief note on how these were selected (e.g., matching VGGT's multi-scale tap points) would be useful.
Appendix C: training uses 64 NVIDIA H20 GPUs for approximately six days. The batch size and effective batch size are not specified; please add these.
Table III: TRIG's OpenScene AUC@30° (69.5) is lower than DVGT's (74.7). The paper does not discuss this regression. A brief acknowledgment and possible explanation would improve the honesty of reporting.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review. The two major comments both identify legitimate issues that require revision: (1) numerical discrepancies between ablation tables and main results tables, and (2) the absence of self-run baselines. We address each below.

read point-by-point responses

Referee: Tables IV, VI, and VII report 'Full model (Ours)' values that systematically differ from the 'TRIG (Ours)' entries in the main results Tables I-III for the same datasets and metrics. Concrete examples: KITTI Acc is 0.333 in Table I vs 0.291 in Table VII; Waymo Comp is 1.192 in Table I vs 0.879 in Table VII; DDAD AbsRel is 0.109 in Table II vs 0.099 in Table VII; KITTI AbsRel is 0.046 in Table II vs 0.041 in Table VII. The ablation tables consistently show better numbers than the main results. If the 'full model' in ablations is not the same checkpoint or evaluation protocol as 'TRIG (Ours)' in the main tables, then the ablation deltas may not reflect the actual contribution of each component to the reported SOTA numbers. The paper provides no explanation.

Authors: The referee is correct that there are numerical discrepancies between the ablation tables and the main results tables, and we acknowledge that the manuscript fails to explain these differences. The root cause is as follows. The main results in Tables I-III are computed on a per-dataset basis using the full evaluation protocol described in the Appendix, including all sequences and all cameras for each benchmark. The ablation tables (IV, VI, VII) use a different evaluation subset: to keep ablation experiments tractable, we evaluated on a fixed subsample of sequences per dataset (approximately 25% of the evaluation set, uniformly sampled). This was done to manage compute cost given the large number of ablation variants (5 variants x 4 datasets x 3 tasks). The subsampling means that the ablation 'Full model' numbers are not directly comparable to the main table numbers, and the ablation deltas reflect component contributions on the subsampled set rather than the full evaluation set. We agree this is a significant presentation problem. In the revision, we will: (a) add an explicit note in the ablation section stating that ablation tables use a subsampled evaluation set and are not directly comparable to the main results, (b) report the full-evaluation-set numbers for the 'Full model' row in each ablation table so readers can verify consistency with the main tables, and (c) for at least Table VII (the dataset-specific ablation), re-run the full model row on the complete evaluation set for all four datasets to eliminate the discrepancy. We acknowledge that the current presentation could mislead readers into thinking the ablation deltas apply to the SOTA numbers directly, and we will fix this. revision: yes
Referee: The statement that 'baseline results are taken from the DVGT paper' means TRIG's numbers are compared against published values from another paper rather than re-implemented baselines under identical training and evaluation conditions. This is load-bearing because the reported improvements are very large. If there are differences in evaluation protocol, input resolution, data preprocessing, or metric computation between the DVGT paper's setup and TRIG's setup, the magnitude of improvement could be inflated. The paper should either re-run at least the strongest baseline (DVGT) under its own pipeline or explicitly confirm that all evaluation settings are identical and document what those settings are.

Authors: This is a fair and important concern. We will address it in two ways in the revision. First, we will re-run the DVGT baseline under our own evaluation pipeline. DVGT's code and pretrained weights are publicly available, so we can run inference on all five benchmarks using our evaluation code, metric implementations, and data preprocessing. This will produce a directly comparable set of DVGT numbers. We will report these self-run DVGT results alongside the published DVGT numbers in the main tables, clearly labeling which is which. Second, we will add a detailed evaluation protocol table to the Appendix documenting all settings that affect comparability: input resolution (518x294 for all methods), evaluation sequence sampling rate (2 Hz), number of frames per sequence, number of cameras per dataset, metric computation code, and whether Sim(3) alignment is applied. We note that our preprocessing already follows the DVGT protocol (as stated in Section V and the Appendix: images resized to 518x294, 2 Hz sampling, same dataset splits), so we expect the self-run numbers to be close to the published ones, but we agree this must be verified empirically rather than assumed. If discrepancies between published and self-run DVGT numbers arise, we will report and discuss them transparently. revision: yes

Circularity Check

0 steps flagged

No significant circularity: trajectory-rig factorization is a standard SE(3) decomposition, priors come from calibrated extrinsics, and supervision uses dataset ground-truth poses.

full rationale

The paper's core derivation chain is self-contained and does not reduce to its inputs by construction. Eq. (2) factorizes a camera pose P^t_{w←c} = P^t_{w←ego} · P^c_{ego←c}, which is a standard SE(3) decomposition of a world-to-camera transform into ego-trajectory and rig extrinsics — not a circular definition. The pose priors (Eq. 3) are encoded from calibrated camera extrinsics provided by the datasets (standard inputs in autonomous driving), not from the model's own outputs. The additive embedding composition (Eq. 5) z^geo = z_traj + z_rig is a modeling choice for token injection, not a definitional identity that would make predictions equivalent to inputs. The decoupled supervision (Eqs. 8–9) computes relative pose losses against ground-truth annotations from dataset labels, partitioned into rig (cross-camera) and trajectory (cross-time) pairs — this is standard supervised learning, not a fitted parameter renamed as prediction. The sparse attention mechanism (Eq. 7) is an architectural design with no circular dependency. No 'prediction' or 'first-principles result' is equivalent to its inputs by construction. The ablation table discrepancies flagged by the skeptic (Table VII vs Tables I–III) are an internal consistency concern for correctness/reproducibility risk, not a circularity issue — they do not indicate that any claimed derivation reduces to a fit or self-citation. The only minor concern is that baseline numbers are taken from the DVGT paper rather than re-implemented, but this is a fairness issue, not circularity. Score 1 reflects the absence of circular derivation chains.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 1 invented entities

The axiom ledger captures the key modeling choices and parameters. The additive embedding composition (Eq. 5) is the most ad-hoc choice; the rig block layer indices are hand-selected without systematic search.

free parameters (3)

Rig block layer indices
Chosen as {0,4,11,17,23} out of 24 layers; no systematic search or justification provided for this specific set vs alternatives (ablation only tests 3-layer variant).
Adapter MLP dimensions
Lightweight adapter projecting geometry features to camera token space; architecture details not fully specified.
Pose encoding MLP dimensions
Two linear projections mapping poses to 1024-dim latent; layer widths not specified beyond output dimension.

axioms (3)

standard math Camera pose factorizes as P_w←c = P_w←ego * P_ego←c (Eq. 2)
Standard SE(3) composition; the ego pose and rig extrinsic are well-defined kinematic quantities.
ad hoc to paper Additive composition of trajectory and rig embeddings captures geometric structure (Eq. 5)
z_geo = z_traj + z_rig assumes the two factors are linearly separable in the learned embedding space; no theoretical justification provided for additivity over alternatives (e.g., concatenation, multiplicative gating).
domain assumption VGGT's DPT and camera heads are sufficient for decoding the adapted features
The paper keeps VGGT's prediction heads unchanged (§IV.A), assuming they generalize to the decoupled representation without modification.

invented entities (1)

Rig Block / Traj Block attention patterns independent evidence
purpose: Separate cross-camera spatial interaction from temporal feature aggregation
Ablation studies (Tables IV, VII) show that removing or modifying rig blocks degrades performance, providing falsifiable evidence within the paper.

pith-pipeline@v1.1.0-glm · 16256 in / 3603 out tokens · 276966 ms · 2026-07-08T23:35:52.048439+00:00 · methodology

0 comments

read the original abstract

Vision-centric autonomous driving requires accurate metric geometry and ego-motion estimation from synchronized multi-camera observations. Recent visual geometry models show strong performance in pose estimation, depth prediction, and 3D reconstruction, but are not tailored to rigid multi-camera driving systems. They often encode camera poses as entangled representations, in which time-varying ego-motion and static camera-rig geometry are jointly modeled, limiting the utilization of vehicle-side geometric priors. We propose Trajectory-Rig Decoupled Metric Geometry Learning (TRIG), a geometry perception framework for autonomous driving. TRIG factorizes camera poses into ego-trajectory and camera-rig components, enabling separate modeling of ego-motion and static multi-camera topology. We introduce decoupled pose encoding and supervision, which separately constrain trajectory evolution and rig geometry for metric-consistent learning. Moreover, sparse Temporal--Spatial attention separates cross-camera interaction from temporal aggregation, reducing global attention cost while preserving geometric reasoning. Experiments on five autonomous driving benchmarks show that TRIG achieves state-of-the-art performance in pose estimation, metric depth prediction, and 3D reconstruction.