Texture Generation Using Dual-Domain Feature Flow with Multi-View Hallucinations
Authors: Seunggyu Chang, Jungchan Cho, Songhwai Oh203-211
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show the superiority of our model for estimating a directly render-able texture map, which is applicable to 3D animation rendering. Furthermore, our model also improves an overall generation quality in the image domain for pose and viewpoint transfer tasks.To validate our method, we conduct experiments on various datasets and show the superiority of our method for generating a texture map and pose transferred images. We also demonstrate our resulting texture map can be applied to 3D human model for rendering a 3D animation clip. We use three datasets to evaluate our model: Deep Fahsion In-shop Clothes Retrieval Benchmark (Liu et al. 2016), i PER (Liu et al. 2019), and Fashion video collected from Amazon (Zablotskaia et al. 2019). We compare our method to several state-of-the-art pose guided image transfer methods including PG2 (Ma et al. 2017), Def-GAN (Siarohin et al. 2018), GFLA (Ren et al. 2020), LWG (Liu et al. 2019), and a recent texture map estimation method, HPBTT (Zhao et al. 2020). The results are summarized in Table 1. To analyze whether the proposed hallucination generation scheme is indeed helpful for texture map estimation, we conduct experiments increasing the number of hallucination, nh, from one to four. Table 2 summarizes the results. To analyze the role of inter-domain feature flows, we conduct ablation studies by unlinking each attention path from one domain to the other. |
| Researcher Affiliation | Academia | Seunggyu Chang1, Jungchan Cho2, Songhwai Oh1 1Department of ECE, ASRI, Seoul National University 2School of Computing, Gachon University seunggyu.chang@rllab.snu.ac.kr, thinkai@gachon.ac.kr, songhwai@snu.ac.kr |
| Pseudocode | No | The paper describes the method using mathematical formulations and descriptive text, but it does not include any formal pseudocode blocks or algorithms labeled as such. |
| Open Source Code | No | The paper does not provide a direct link to a code repository or an explicit statement about releasing the source code for the described methodology. |
| Open Datasets | Yes | We use three datasets to evaluate our model: Deep Fahsion In-shop Clothes Retrieval Benchmark (Liu et al. 2016), i PER (Liu et al. 2019), and Fashion video collected from Amazon (Zablotskaia et al. 2019). |
| Dataset Splits | No | The paper mentions 'training set' (e.g., 'From Deep Fahsion we filter out 5,745 images ... from the training set.') and discusses aspects of training strategy (e.g., 'most training images are biased to frontal and side views'), but it does not specify explicit train/validation/test splits by percentage or sample counts needed for reproduction. |
| Hardware Specification | No | The paper does not explicitly describe any specific hardware components (e.g., GPU models, CPU types, memory specifications) used to run the experiments. |
| Software Dependencies | No | The paper mentions using 'pretrained VGG19 networks', 'SMPL (Loper et al. 2015)', and 'Open Pose (Cao et al. 2019)', but it does not specify version numbers for any of these software components or other libraries used for implementation, which is required for reproducible description. |
| Experiment Setup | No | The paper includes a 'Loss Functions and Training Strategy' section, which describes the general training process and loss functions used. However, it does not provide specific hyperparameters such as learning rate, batch size, number of epochs, or optimizer settings, which are crucial for replicating the experimental setup. |