Composite image degradations, where rain, haze, low-light, and snow overlap spatially within a single frame, pose a fundamental challenge for unified restoration models. We identify two failure modes on such inputs. First, deterministic flow transport averages over feasible solutions and collapses to over-smoothed predictions. Second, image-level expert routing yields a selectivity ratio of only 1.2 × , with near-uniform weights that leave experts under-specialized on localized distortions. Motivated by these observations, we propose F 2 D-Net, a factorized stochastic transport framework. The model factorizes the restoration velocity field into a shared backbone and degradation-specific expert increments. A learned spatial gating then assembles them at every pixel. The core mechanism is a state-dependent multiplicative noise whose magnitude scales with the local residual. The noise remains large far from the clean target to encourage diverse reconstructions, and contracts to zero at convergence to preserve fine detail. The closed-form log-normal transition collapses sampling to one network call, matching feed-forward speed. On CDD-11, F 2 D-Net attains the best structural and perceptual quality across all complexity tiers with substantially higher expert selectivity than image-level routing. It also remains competitive on standard three-task and five-task all-in-one benchmarks. Our project page is at https://sxvvv.github.io/f2net .

Visible light positioning (VLP) has become one of the most important technologies for providing highly accurate location-based services. Despite this, line-of-sight (LOS) VLP systems have severe limitations when operating in complex and dynamic indoor environments. We propose a non-LOS VLP system based on an image sensor (IS) and pixel coordinate fingerprinting operating under shadowing and blocking. The system does not require communication between the transmitter and the receiver eliminating the need for a complex modulation process. For the first time, we utilize the highlighted center pixel coordinates on an IS as the fingerprint features and extract highlight center coordinates using the deep learning model Yolov8. Furthermore, we propose an Elastic Net regression with Weighted K-Nearest Neighbor Residual Correction algorithm to improve the positioning performance, which employs an Elastic Net for global prediction and a weight k-nearest neighbor algorithm to correct the global prediction residuals by leveraging the residual information from weighted neighboring samples. The experimental results show that the average positioning error and 90th percentile errors of the proposed system are 4.77 and 6.67 cm, respectively, with only 16 training points. In addition, the stability of the system is verified by arbitrary and diagonal sets.

Visible light positioning (VLP) is a low-cost, highly accurate alternative localization technology for indoor applications that makes use of existing light emitting diode (LED)-based lights, which is highly accurate and low costs. It is, however, a major challenge for the existing VLP systems to achieve line-of-sight (LOS) positioning in complex and variable indoor environments. We propose a non-LOS (NLOS) fingerprint-based VLP system based on a multipixel photon counter (MPPC) to address the problem of obstructed LOS paths. Using MPPC, very faint light can be detected with a very high sensitivity and excellent photon counting capability, which enhances the ability to recognize and detect signals in an NLOS environment. We propose a novel method of generating fingerprint database using the NLOS channel model, which construct the relationship between the received signal strength and the distance from MPPC to the virtual image of LED interpolated by only knowing the distance between the LED and the interpolation position. Furthermore, we propose an optimal parameter weighted K-nearest neighbor algorithm, which utilizes the mean absolute error (MAE) as the evaluation metric. In this algorithm, a grid search method is employed to determine the optimal number of neighbors and the distance metric for each test point, thereby enhancing the positioning accuracy. Using only 25 offline measurements, the measured average positioning error (PE) and 90th percentile error are 4.02 and 9.98 cm, respectively, when the MPPC height is 70 cm.

Unsupervised Domain Adaptation (UDA) refers to the method that utilizes annotated source domain data and unlabeled target domain data to train a model capable of generalizing to the target domain data. Domain discrepancy leads to a significant decrease in the performance of general network models trained on the source domain data when applied to the target domain. We introduce a straightforward approach to mitigate the domain discrepancy, which necessitates no additional parameter calculations and seamlessly integrates with self-training-based UDA methods. Through the transfer of the target domain style to the source domain in the latent feature space, the model is trained to prioritize the target domain style during the decision-making process. We tackle the problem at both the image-level and shallow feature map level by transferring the style information from the target domain to the source domain data. As a result, we obtain a model that exhibits superior performance on the target domain. Our method yields remarkable enhancements in the state-of-the-art performance for synthetic-to-real UDA tasks. For example, our proposed method attains a noteworthy UDA performance of 76.93 mIoU on the GTA→Cityscapes dataset, representing a notable improvement of +1.03 percentage points over the previous state-of-the-art results.

Visible light positioning (VLP) is one of the effective solutions for low-cost and high-precision indoor positioning due to its advantages of low cost, high energy efficiency, and immunity from radio frequency induced electromagnetic interference. However, severe performance limitations in occluded environments present a major challenge for the development of VLP. We propose a non-line-of-sight (NLOS) VLP system based on a multipixel photon counter (MPPC) to solve the problem of obstructed line-of-sight (LOS) paths. MPPC consists of multiple avalanche photodiodes operating in Geiger mode, which has very high sensitivity, excellent photon counting capability and can perform detection of very weak light at photon counting levels. It can effectively improve the recognition and detection of signals in NLOS environments. Using the NLOS channel model, we innovatively propose a NLOS-MPPC-VLP positioning theory, which utilizes the NLOS optical signals received by the MPPC to construct the relationship between the received signal strength and the distance from the MPPC to the virtual images of the light emitting diode (LED). Through this theory, the trilateration algorithm is used to calculate the position of the receiver. The experimental results show that the average positioning errors for the proposed system are 8.77 and 10.15 cm for 70 and 100-cm MPPC heights, respectively.

In many mobile visual analysis applications, compressed video is transmitted over a communication network and analyzed by a server. Typical processing steps performed at the server include keypoint detection, descriptor calculation, and feature matching. Video compression has been shown to have an adverse effect on feature-matching performance. The negative impact of compression can be reduced by using the keypoints extracted from the uncompressed video to calculate descriptors from the compressed video. Based on this observation, we propose to provide these keypoints to the server as side information and to extract only the descriptors from the compressed video. First, we introduce four different frame types for keypoint encoding to address different types of changes in video content. These frame types represent a new scene, the same scene, a slowly changing scene, or a rapidly moving scene, and are determined by comparing features between successive video frames. Then, we propose Intra, Skip, and Inter modes of encoding the keypoints for different frame types. For example, keypoints for new scenes are encoded using the Intra mode, and keypoints for unchanged scenes are skipped. As a result, the bitrate of the side information related to keypoint encoding is significantly reduced. Finally, we present pairwise matching and image retrieval experiments conducted to evaluate the performance of the proposed approach using the Stanford mobile augmented reality dataset and 720p format videos. The results show that the proposed approach offers significantly improved feature matching and image retrieval performance at a given bitrate.

In many mobile visual analysis scenarios, compressed images are transmitted over a communication network for analysis at a server. Often, the processing at the server includes some form of feature extraction and matching. Image compression has been shown to have an adverse effect on feature matching performance. To address this issue, we propose to signal the feature keypoints as side information to the server, and extract only the feature descriptors from the compressed images. To this end, we propose an approach to efficiently encode the locations, scales, and orientations of keypoints extracted from the original image. Furthermore, we propose a new approach for selecting relevant yet fragile keypoints as side information for the image, thus further reducing the data volume. We evaluate the performance of our approach using the Stanford mobile augmented reality dataset. Results show that the feature matching performance is significantly improved for images at low bitrate.

This paper presents a novel rate control framework for H.264/Advanced Video Coding-based video coding that improves the preservation of gradient-based features like scale-invariant feature transform or speeded up robust feature compared with the default rate control algorithm in the JM reference software. First, a criterion (matching score) for feature preservation on the basis of the bag-of-features concept is proposed. Then, the matching scores are collected as a function of the quantization parameters and analyzed for different feature types. With this analysis, macroblocks are categorized into different groups before encoding. Our rate control algorithm assigns different quantization parameters to each group according to the importance of the group for feature extraction. The experimental results show that our rate control algorithm achieves the desired target bit rate, and more features are preserved compared with videos encoded using the default rate control. The proposed approach not only improves feature preservation, but also leads to a noticeable performance improvement in a real image retrieval system. The rate control framework proposed in this paper is fully standard compatible.

Keypoint or interest point detection is the first step in many computer vision algorithms. The detection performance of the state-of-the-art detectors is, however, strongly influenced by compression artifacts, especially at low bit rates. In this paper, we design a novel quantization table for the widely-used JPEG compression standard which leads to improved feature detection performance. After analyzing several popular scale-space based detectors, we propose a novel quantization table which is based on the observed impact of scale-space processing on the DCT basis functions. Experimental results show that the novel quantization table outperforms the JPEG default quantization table in terms of feature repeatability, number of correspondences, matching score, and number of correct matches.

We study the impact of JPEG compression on the performance of an image retrieval system for different feature detector-descriptor combinations. The VLBenchmarks retrieval framework is used to compare a total of 60 detector-descriptor combinations for a dataset with JPEG-encoded query images. Our results show that among all tested detectors, the Hessian-Affine detector leads to the most robust performance in the presence of strong JPEG compression. Additionally, we compare the retrieval gains of the different detector-descriptor pairs after processing the JPEG-encoded query images with different deblocking filters. The results illustrate that for the MSER, MFD and WαSH detectors, the retrieval results benefit from two of the deblocking approaches at low bit rate irrespective of what descriptor the detectors are combined with. The same two deblocking filters are found to increase the retrieval performance for the MROGH descriptor when combined with most of the tested detectors.

Compression artifacts in low-quality videos strongly influence the performance of feature matching algorithms. In order to achieve reasonable feature matching performance even for low bit rate video, we propose to allocate the bit budget during compression such that the important features are preserved. Specifically, we present two bit allocation approaches to preserve the strongest SIFT features for H.264 encoded videos. For both approaches, we first categorize the Macroblocks in a Group of Pictures into several groups according to the scale specific characteristics of SIFT features. In our first approach a novel R-D model based on the matching score is applied to allocate the bit budget to these groups. In our second approach, in order to reduce the computational complexity, we analyze the detector characteristics of correctly matched pairs and propose a R-D optimization method based on the repeatability metric. Our experiments show that both approaches achieve better feature preservation when compared to standard video encoding which is optimized for maximum picture quality. The proposed approaches are fully standard compatible and the encoded videos can be decoded by any H.264 decoder.

For image compression applications where the information sink is not a person but a computer algorithm, the image encoder should control the encoding process in such a way that the important and relevant features of the image are preserved after compression. In this paper, our goal is to preserve the strongest SIFT features for JPEG-encoded images. We analyze the relevant characteristics of SIFT features and categorize the image Macroblocks into several groups. Then we propose a novel rate-distortion model which is based on the SIFT feature matching score. The dependency between the quantization table in the JPEG file and the common Lagrange multiplier is obtained from a training image database. Then for a given image quality we exploit this relationship to perform R-D optimization for each group. Our results show that the proposed algorithm achieves better feature preservation when compared to standard JPEG encoding. The proposed approach is fully standard compatible.