InstantSticker: Realistic Decal Blending via Disentangled Object Reconstruction

Authors: Yi Zhang, Xiaoyang Huang, Yishun Dou, Yue Shi, Rui Shi, Ye Chen, Bingbing Ni, Wenjun Zhang

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiment, we introduce the Ratio Variance Warping (RVW) metric to evaluate the local geometric warping of the decal area. Extensive experimental results demonstrate that our method surpasses previous decal blending methods in terms of editing quality, editing speed and rendering speed, achieving the state-of-the-art. We evaluate our method on Ne RF-Synthetic (Mildenhall et al. 2020), Shiny Blender (Verbin et al. 2022), and DTU (Jensen et al. 2014) dataset. We mainly compare our method with the following methods capable of performing decal blending on reconstructed objects: Neu Mesh, Seal-3D, and DE-Ne RF. We present the components of disentangling and editing results in Fig. 5. Since disentangled reconstruction is more challenging than free-form reconstruction, we compare our approach with advanced disentangled reconstruction methods for fairness, in terms of PSNR. The results demonstrate that our reconstruction quality is comparable to that of DE-Ne RF, as shown in Tab. 1. In this section, we evaluate several technical components in our pipeline through a series of ablation studies.
Researcher Affiliation Collaboration Yi Zhang1, Xiaoyang Huang1, Yishun Dou2, Yue Shi1, Rui Shi1, Ye Chen1, Bingbing Ni1*, Wenjun Zhang1 1Shanghai Jiao Tong University, Shanghai 200240, China 2Huawei EMAIL
Pseudocode No The paper describes the method and pipeline in prose and diagrams (Fig. 2) but does not provide structured pseudocode or algorithm blocks.
Open Source Code No The paper does not explicitly state that the source code for their methodology is released, nor does it provide a link to a code repository.
Open Datasets Yes We evaluate our method on Ne RF-Synthetic (Mildenhall et al. 2020), Shiny Blender (Verbin et al. 2022), and DTU (Jensen et al. 2014) dataset.
Dataset Splits No The paper mentions using Ne RF-Synthetic, Shiny Blender, and DTU datasets but does not provide specific details on how these datasets were split into training, validation, or test sets.
Hardware Specification Yes Training is performed on a single NVIDIA RTX 3090 GPU and takes only 1 hour.
Software Dependencies No The paper mentions using Py Torch3D, Open GL, and Libigl library but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes The Adam optimizer is used, with a learning rate of 5e-4 for Ft and Fr, and 5e-3 for the appearance features and environment map. Training is performed on a single NVIDIA RTX 3090 GPU and takes only 1 hour. The positional encoding dimension is set to 4, and the hidden layer dimension of MLP is 256.