AWRaCLe: All-Weather Image Restoration Using Visual In-Context Learning
Authors: Sudarshan Rajagopalan, Vishal M. Patel
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we demonstrate the effectiveness of AWRa CLe for all-weather restoration and show that our method advances the state-of-the-art in AWIR. ... 4 Experimental Results In this section, we explain our implementation, datasets used, results and ablation studies. ... Table 1: Quantitative comparisons of AWRa CLe with SOTA on the test sets described in Sec. 4.2. |
| Researcher Affiliation | Academia | Sudarshan Rajagopalan, Vishal M. Patel Johns Hopkins University EMAIL |
| Pseudocode | No | The paper describes the methodology with equations and block diagrams (Figure 2) and detailed explanations of DCE and CF blocks but does not include a distinct pseudocode or algorithm block. |
| Open Source Code | No | Project Page https://sudraj2002.github.io/awraclepage/. The paper provides a project page URL, which is a demonstration page and not a direct link to a code repository. There is no explicit statement about releasing the source code. |
| Open Datasets | Yes | We use the Snow100k (Liu et al. 2018), synthetic rain (Zamir et al. 2021) datasets (SRD) and RESIDE (Li et al. 2019) to train our method for all-weather restoration. ... For deraining, we evaluate the methods on Rain100H (Yang et al. 2017) for heavy rain and Rain100L (Yang et al. 2017) for light rain... For dehazing, we use RESIDE s Synthetic Objective Testing Set (SOTS) outdoor... Foggy Cityscapes (Hahner et al. 2019) for dehazing, rain images from Rain DS (Quan et al. 2021) for deraining, and Snow Cityscapes (Zhang et al. 2021b) for desnowing. |
| Dataset Splits | Yes | The training split of Snow100k contains 50, 000 synthetic snow images along with the corresponding clean images. ... For deraining, we use the training split of SRD containing 13, 711 clean-synthetic rainy image pairs. For dehazing, we use the Outdoor Training Set (OTS) of RESIDE which consists of 72, 135 clean-synthetic hazy image pairs for training. ... In summary, we obtain 12, 077 light rain images, 1, 634 heavy rain images, 38, 921 light haze images, 33, 214 heavy haze images, 37, 122 light snow images and 12, 878 heavy snow images for training. ... For desnowing, we use the test split of Snow100k dataset containing 50, 000 paired images. For deraining, we evaluate the methods on Rain100H (Yang et al. 2017) for heavy rain and Rain100L (Yang et al. 2017) for light rain, each consisting of 100 paired images. For dehazing, we use RESIDE s Synthetic Objective Testing Set (SOTS) outdoor containing 500 paired images. |
| Hardware Specification | Yes | Our method is trained using the Adam W optimizer with a cosine annealing Learning Rate (LR) scheduler. We train for a total of 100 epochs on 8 RTX A5000 GPUs with a batch size of 32, initial LR= 2 10 4, weight decay= 0.01, β1 = 0.9, β2 = 0.999 and warm-up for 15 epochs. |
| Software Dependencies | No | Our implementation utilized Py Torch (Paszke et al. 2019). The paper mentions PyTorch but does not provide a specific version number, which is required for reproducibility. |
| Experiment Setup | Yes | Our method is trained using the Adam W optimizer with a cosine annealing Learning Rate (LR) scheduler. We train for a total of 100 epochs on 8 RTX A5000 GPUs with a batch size of 32, initial LR= 2 10 4, weight decay= 0.01, β1 = 0.9, β2 = 0.999 and warm-up for 15 epochs. We use random crop size of 128 128 pixels, and random flipping as data augmentations. The loss function used is the L1 loss. For extracting CLIP features, no augmentations are used and the images in the context pair are resized to 224 224. |