- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents DACA, a zero-shot 3D visual grounding framework built on a VLM and visual tools. It extends TAB with optical-flow-guided temporal tracking, multi-keypoint geometric projection, and semantic-geometric cross-validation. The goal is to collect reliable object masks across views and reconstruct the target’s 3D bounding box. Experiments on 250-example subsets of ScanRefer and Nr3D show improvements over TAB, with ablations on the main components.
Critical strengths
-
The motivation is reasonable. Semantic tracking can drift, while geometric projection can select the wrong object. Checking the two against each other is a useful way to filter unreliable observations.
-
The results improve over TAB on both benchmarks using the same reported VLM backbone. The cross-validation ablations also show consistent gains. This provides some evidence that the proposed checks help beyond simply adding more views.
Critical weaknesses
-
The evaluation is too limited to support the main claims. Each benchmark uses only 250 examples. The 2.4 percentage-point gain over TAB corresponds to six additional correct predictions on each dataset. No repeated runs or uncertainty estimates are reported. Both benchmarks also come from ScanNet, so they provide limited evidence of generalization. The paper uses corrected annotations from TAB, but it is unclear whether all baseline numbers use the same examples and annotations. This also makes the claimed advantage over supervised methods hard to assess.
-
The ablations do not fully separate the proposed mechanism from extra processing. Table 3(a) mainly compares against a single-frame baseline. The cross-validation ablation is useful, but there is no comparison that matches the number of views, tool calls, and iterations while removing adaptive feedback. There is also no direct ablation of TSDF visibility against the previous depth-based check, or flow-guided tracking against VLM tracking under the same view budget. The current results do not show how much of the gain comes from the claimed collaboration.
-
Several technical details are inconsistent or insufficiently defined. Eq. (2) uses forward-flow coordinates to sample the reference mask, which appears inconsistent with backward sampling in grid_sample. Eq. (5) compares image-plane coordinates, but the corresponding threshold is reported as 0.4 meters. The definitions of view coverage and reconstruction convergence are also missing. These details affect tracking, validation, and termination, so they matter for reproducing the method.
Questions for authors
-
How were the 250 examples selected for each benchmark? Did all methods use the same examples and corrected annotations? Which baselines did you rerun, and which results were taken from prior work?
-
Are the gains over TAB consistent across repeated runs and different evaluation subsets?
-
Does iterative feedback improve results beyond simply adding more views? Please compare with a fixed pipeline using the same number of views and tool calls, and report accuracy after each iteration.