Thanks for your helpful review
Many thanks for your valuable reviews. Here are our responses.
**[1. Number of scenes to evaluate]**
We follow the evaluation settings used by PlanarRecon and Atlas. Our evaluation is conducted on the official validation set of ScanNet v2, which consists of 312 scenes.
**[2. Code release]**
Yes, our code will be released once the paper is public.
**[3. Discussions of primitive-based optimization's advantages and Airplanes]**
We appreciate the reviewer mentioning AirPlanes, a concurrent great work (not publicly visible when this paper is submitted) presented at CVPR 2024. While both approaches aim for 3D planar reconstruction task, our method offers several distinct advantages:
1. Explicit geometric structure preservation: Compared to point/voxel-based dense reconstruction, using plane/primitive as the basic units explicitly enforces the local geometric (planar) structure during the reconstruction process. Through optimization and merging, these explicit geometric constraints are extended to the entire planar structure. This geometric structure preservation is not trivial or even challenging for depth map / dense points representation, For example, the dense 3D reconstruction of a wall often contains local non-planar / non-smooth regions and noisy outlier points, which affect the subsequent plane label fitting and the resulting geometric accuracy, since RANSAC only responses for extracting the highest-scoring planes with no further optimizations. To the contrary, our planar primitive (AlphaTablets) based optimization pipeline is more straightforward and arguably more effective for the planar reconstruction task, with as-planar-as-possible guaranteed and optimization in the loop.
2. Generic plane representation: We proposed AlphaTablets as a general 3D plane representation, with a 3D planar reconstruction pipeline demonstrating its effectiveness. In contrast, AirPlanes mainly focus on the modular design of 3D planar reconstruction system, where planes are represented with point groups. Our explicit continuous 3D plane representation enables advanced applications such as texture editing and novel view synthesis. Furthermore, if needed, the resulting 3D planar reconstruction of AirPlanes could be represented as AlphaTablets to enable further optimization w.r.t input images for both plane geometry and textures, which is not easy for point or mesh representations (optimization could break the planar structure).
3. Better generalization without dataset-specific training: Our method can operate without the need for training on specific datasets, leading to better generalization across various datasets. In contrast, AirPlanes needs to be trained (especially the plane embedding) with datasets, which may limit its applicability to diverse scenarios without further fine-tuning.
4. Complementary contributions: The main contribution of AirPlanes, embedding grouping, is orthogonal to our approach. Planar embeddings could serve as a similarity measurement and be integrated into our merging stage, suggesting exciting possibilities for future research.