Cross-Image Context for Single Image Inpainting
Authors: Tingliang Feng, Wei Feng, Weiqi Li, Di Lin
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results demonstrate the effectiveness and generalization of CICM, which achieves state-of-the-art performances on various datasets for single image inpainting. We intensively evaluate CICM on the datasets (i.e., Places2 [9] and Celeb A [10]) for image inpainting, where we achieve the state-of-the-art performances. |
| Researcher Affiliation | Academia | Tingliang Feng, Wei Feng, Weiqi Li, Di Lin College of Intelligence and Computing, Tianjin University {fengtl, wicky}@tju.edu.cn, wfeng@ieee.org, Ande.lin1988@gmail.com |
| Pseudocode | No | No figure or section explicitly labeled "Pseudocode" or "Algorithm" was found. |
| Open Source Code | Yes | Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] |
| Open Datasets | Yes | We evaluate our approach on the Places2 [9] and Celeb A [10] datasets. Places2 provides over 8M images taken from over 365 scenes for training, along with 30K images for testing. It is used for the internal study of our approach. We also compare our approach with state-of-the-art methods on the Celeb A dataset, which provides about 163K and 20K images, respectively, for training and testing. We use the Cityscapes [11] and Outdoor Scenes [12] to extensively justify the effectiveness of our method. |
| Dataset Splits | No | Places2 provides over 8M images taken from over 365 scenes for training, along with 30K images for testing. ... Celeb A dataset, which provides about 163K and 20K images, respectively, for training and testing. ... Cityscapes contains 5,000 street-view images, where 2,975 and 1,525 images are used for training and testing. Outdoor Scenes contains 9,900 training and 300 testing images. The paper specifies train/test splits, but does not explicitly mention a separate validation set split or how it was derived. |
| Hardware Specification | No | The paper states in its checklist that it included "the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)", marked as [Yes], but these specific hardware details (e.g., GPU model, CPU type) are not explicitly present in the provided paper text. |
| Software Dependencies | No | The paper does not list any specific software dependencies with version numbers (e.g., 'PyTorch 1.9', 'Python 3.8', 'CUDA 11.1'). |
| Experiment Setup | Yes | During the network training, we resort to L2-norm and the adversarial loss [44] for measuring the inpainting error as: Linpaint = ||I I||2, Ladv = EI log(1 D(I, I )) EI log(D(I , I)) , (5) where D(x, y) = sigmoid(C(x) Ey [C(y)]). (6) In Eq.(5), I RH W 3 is the ground-truth result. Ex is the expectation with respect to x. In Eq.(6), C is a discriminator network. We also use L2-norm for measuring the estimated corrupted ratio and the ground-truth ratio of the image region as: n=1 ||γn γn||2, (7) where γn and γn are the estimated and ground-truth corrupted ratios for the nth region in the input image I. Lratio is minimized. ... L = λ1Linpaint + λ2Ladv + λ3Lratio + λ4Linter λ5Lintra, (9) where λ1 = 1.0, λ2 = 0.1, λ3 = 1.0, λ4 = 20 and λ5 = 0.5. We minimize the overall objective L to optimize the network parameters. ... We also use the momentum factor β = 0.5 to control the update of the anchor feature. |