- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes a new task, 3D Long-Term Changing Grounding, a variant of 3D visual grounding in which object layouts evolve across multiple timestamps and the model is expected to localize referred objects while limiting the cost of repeated scene perception. To study this problem, the authors construct TemScan, a synthetic benchmark based on ProcTHOR/AI2-THOR. The authors further develop a detect-then-ground framework for this task, and experimental results show that the proposed method achieves competitive performance on the new benchmark.
Critical strengths
-
The paper identifies a meaningful limitation of current 3D visual grounding benchmarks. Existing benchmarks largely assume a fixed scene, where an embodied system must reason from scene information that becomes stale over time. Introducing temporal scene changes together with an explicit perception-cost dimension is therefore a useful direction.
-
The proposed system explicitly considers both final accuracy and inference cost, which is essential for practical application.
Critical weaknesses
-
The evaluation protocol contains important ambiguities regarding test time information and possible test set leakage. The paper states that the LLM inference experiments are conducted on the 30 test scenes in the appendix, while the skill mechanism is constructed by mining successful and failed reasoning trajectories. It is not specified whether these trajectories are obtained exclusively from the training/validation split. If test trajectories are used to select frequent tool flows or generate the skill descriptions, the reported test results would no longer constitute a clean evaluation. In addition, the appendix shows that the agent receives pre-parsed target, anchor, and relation fields in addition to the raw referring expression, but the paper does not explain how these fields are produced or whether competing methods receive equivalent structured information. If these fields come directly from dataset annotations, they provide substantial oracle information at test time.
-
The experiments do not sufficiently isolate the contribution of the proposed uncertainty-aware temporal memory and do not provide a fully controlled comparison. The uncertainty score is essentially the elapsed time since an object/room was updated, and the method scans the room with the highest average score. Therefore, it is unclear whether the reported improvement comes from the proposed uncertainty mechanism or simply from periodically refreshing stale rooms. The ablation compares the full system mainly against retaining the original memory, rather than equal budget alternatives such as random, query-conditioned, or periodic rescanning. Moreover, the proposed method uses DeepSeek-Chat while the major 3DVG baselines in Table 2 use GPT-4o, and the closest dynamic-scene grounding method discussed in the related work is not included in the empirical comparison. These factors make it difficult to attribute the reported advantages specifically to the proposed method.
Questions for authors
-
Which data split is used to construct the reusable skills? Please clarify where the successful/failed trajectories used to select the top-10 tool flows and generate the skill descriptions come from.
-
How are the target, anchor, and relation fields shown in the Appendix prompt obtained at test time? Are they parsed automatically from the raw referring expression by the evaluated system, or are they taken from the structured annotations used to generate TemScan queries? Do all compared methods receive the same structured fields? Please also clarify whether scan_room_from_gt / scan_full_scene_from_gt accesses only FCAF3D predictions obtained from a newly observed current timestamp point cloud, or whether any ground truth object/room information is used.
-
How is the persistent scene graph handled across multiple queries from the same timestamp? Since each timestamp contains multiple queries and a room is rescanned before each query, are scene-graph updates accumulated across those queries? And equal budget comparison is important for determining whether the proposed uncertainty policy itself provides benefit.