This paper reviews the NTIRE 2022 challenge on night photography rendering. The challenge solicited solutions that processed RAW camera images captured in night scenes to produce a photo-finished output image encoded in the standard RGB (sRGB) space. Given the subjective nature of this task, the proposed solutions were evaluated based on the mean opinions of viewers asked to judge the visual appearance of the results. Michael Freeman, a world-renowned photographer, further ranked the solutions with the highest mean opinion scores. A total of 13 teams competed in the final phase of the challenge. The proposed methods provided by the participating teams represent state-of-the-art performance in nighttime photography. Results from the various teams can be found here: https://nightimaging.org/

NTIRE 2022 Challenge on Night Photography Rendering

In this paper, we present a novel method for generating synthetic underwater images considering revised image formation model. We propose to use the generated synthetic underwater images to train a conditional generative adversarial network (CGAN) towards restoration of degraded underwater images. Restoration of degraded underwater images using traditional dehazing models is challenging as they are insensitive to wavelength, depth, water type and treat backscattering and direct signal attenuation coefficients to be equal. However, learning based models for restoration perform well but sensitive to availability of ground truth information. Generating ground truth labels in underwater scenario demands in-situ measurements using expensive equipments and is infeasible due to varying underwater currents. Towards this, we propose to generate synthetic underwater images using revised image formation model. Revised image formation model is sensitive to different attenuation coefficients: 1) back scattering, 2) direct scattering and 3) veiling light. We propose to estimate these attenuation coefficients considering proven facts from the literature. We demonstrate restoration of real underwater images through restoration framework trained using rendered synthetic underwater images, and compare results of restoration with state-of-the-art techniques.

/pdf/ruig-realistic-underwater-image-generation-towards-56qwe8ez8c.pdf

RUIG: Realistic Underwater Image Generation Towards Restoration

This paper reviews the NTIRE 2022 challenge on night photography rendering. The challenge solicited solutions that processed RAW camera images captured in night scenes to produce a photo-finished output image encoded in the standard RGB (sRGB) space. Given the subjective nature of this task, the proposed solutions were evaluated based on the mean opinions of viewers asked to judge the visual appearance of the results. Michael Freeman, a world-renowned photographer, further ranked the solutions with the highest mean opinion scores. A total of 13 teams competed in the final phase of the challenge. The proposed methods provided by the participating teams represent state-of-the-art performance in nighttime photography. Results from the various teams can be found here: https://nightimaging.org/ 

/pdf/ntire-2022-challenge-on-night-photography-rendering-381bsjxr.pdf

The aim of this study was the colour fastness investigation of the new synthetized direct symmetrical azo-stilbene dye, using two of the film field domain reference tests, namely the wet-scrub and the UV tests. The dye was incorporated in a water-based resin, and then was applied on a PCV foil. The film colour parameters were determined before and after 200 wet-scrub cycles. Further, the dye was mixed with an acrylo-polyurethanic resin and then was applied on an aluminium plate, which was exposed to UV radiation for 414 h. The film colour parameters were recorded periodically. The maxima of the reflectance spectra depend on the UV time exposure. The quality of the film was analysed by the degree of gloss. A second focus was the elimination of the dye’s traces from wastewaters (these may be resulted from the industries which apply surface coating methods), using the active carbon powder. The main operational adsorption process parameters influence were investigated. Equilibrium, kinetic and thermodynamic studies were performed. The adsorption process was confirmed by the CIEL*a*b* colour space analysis. All colour studies were investigated using UV–Vis spectroscopy.

/pdf/aspects-regarding-colour-fastness-and-adsorption-studies-of-35q1dm79rr.pdf

Aspects regarding colour fastness and adsorption studies of a new azo-stilbene dye for acrylic resins.

K-BP neural network-based strain field inversion and load identification for CFRP

Adaptive Cubic Spline Interpolation in CIELAB Color Space for Underwater Image Enhancement

Vision-Language models like CLIP have been widely adopted for various tasks due to their impressive zero-shot capabilities. However, CLIP is not suitable for extracting 3D geometric features as it was trained on only images and text by natural language supervision. We work on addressing this limitation and propose a new framework termed CG3D (CLIP Goes 3D) where a 3D encoder is learned to exhibit zero-shot capabilities. CG3D is trained using triplets of pointclouds, corresponding rendered 2D images, and texts using natural language supervision. To align the features in a multimodal embedding space, we utilize contrastive loss on 3D features obtained from the 3D encoder, as well as visual and text features extracted from CLIP. We note that the natural images used to train CLIP and the rendered 2D images in CG3D have a distribution shift. Attempting to train the visual and text encoder to account for this shift results in catastrophic forgetting and a notable decrease in performance. To solve this, we employ prompt tuning and introduce trainable parameters in the input space to shift CLIP towards the 3D pre-training dataset utilized in CG3D. We extensively test our pre-trained CG3D framework and demonstrate its impressive capabilities in zero-shot, open scene understanding, and retrieval tasks. Further, it also serves as strong starting weights for fine-tuning in downstream 3D recognition tasks.

CLIP goes 3D: Leveraging Prompt Tuning for Language Grounded 3D Recognition

In this paper, we propose a method to refine sparse point clouds of complex structures generated by Structure from Motion in order to achieve improved visual fidelity of ancient Indian heritage sites. We compare our results with the state-of-the-art upsampling networks.

Refining SfM Reconstructed Models of Indian Heritage Sites

Simultaneous Localization and Mapping (SLAM) is a core component in navigation systems of numerous vehicles such as UAVs and field robots. Relocalization in visual SLAM is the identification of camera position in a mapped area. The necessity for such a system arises in case of failed tracking or errors in loop closing. These processes are computationally intensive, and have real-time requirements and must run on processors with power and memory restrictions. We propose a system for relocalization which identifies 6 degrees of freedom camera pose and trajectory based on a single query image by using a vocabulary tree to create a visual word dictionary for quick and efficient image retrieval. This system eliminates the need to build and run computationally heavy complete SLAM systems and does not require storage of large video files/datasets on the device. We demonstrate the proposed system on a low memory embedded system, the ARM Cortex A53 board (Raspberry Pi 3 Model B). The proposed system enables faster, low memory relocalization with improved memory usage and time taken for inference in specialized, low cost systems such as smaller autonomous robots and drones.

Relocalization of Camera in a 3D Map on Memory Restricted Devices

Pseudo-label based self training approaches are a popular method for source-free unsupervised domain adaptation. However, their efficacy depends on the quality of the labels generated by the source trained model. These labels may be incorrect with high confidence, rendering thresholding methods ineffective. In order to avoid reinforcing errors caused by label noise, we propose an uncertainty-aware mean teacher framework which implicitly filters incorrect pseudo-labels during training. Leveraging model uncertainty allows the mean teacher network to perform implicit filtering by down-weighing losses corresponding uncertain pseudo-labels. Effectively, we perform automatic soft-sampling of pseudo-labeled data while aligning predictions from the student and teacher networks. We demonstrate our method on several domain adaptation scenarios, from cross-dataset to cross-weather conditions, and achieve state-of-the-art performance in these cases, on the KITTI lidar target dataset.

Deepti Hegde

Papers

Adaptive Cubic Spline Interpolation in CIELAB Color Space for Underwater Image Enhancement

CLIP goes 3D: Leveraging Prompt Tuning for Language Grounded 3D Recognition

Refining SfM Reconstructed Models of Indian Heritage Sites

Relocalization of Camera in a 3D Map on Memory Restricted Devices

Uncertainty-aware Mean Teacher for Source-free Unsupervised Domain Adaptive 3D Object Detection