Currently, most adverse weather removal tasks are handled independently, such as deraining, desnowing, and dehazing. However, in autonomous driving scenarios, the type, intensity, and mixing degree of the weather are unknown, so the separated task setting cannot deal with these complex conditions well. Besides, the vision applications in autonomous driving often aim at high-level tasks, but existing weather removal methods neglect the connection between performance on perceptual tasks and signal fidelity. To this end, in upstream task, we propose a novel \textbf{Mixture of Weather Experts(MoWE)} Transformer framework to handle complex weather removal in a perception-aware fashion. We design a \textbf{Weather-aware Router} to make the experts targeted more relevant to weather types while without the need for weather type labels during inference. To handle diverse weather conditions, we propose \textbf{Multi-scale Experts} to fuse information among neighbor tokens. In downstream task, we propose a \textbf{Label-free Perception-aware Metric} to measure whether the outputs of image processing models are suitable for high level perception tasks without the demand for semantic labels. We collect a syntactic dataset \textbf{MAW-Sim} towards autonomous driving scenarios to benchmark the multiple weather removal performance of existing methods. Our MoWE achieves SOTA performance in upstream task on the proposed dataset and two public datasets, i.e. All-Weather and Rain/Fog-Cityscapes, and also have better perceptual results in downstream segmentation task compared to other methods. Our codes and datasets will be released after acceptance.
We consider the problem of lossy image compression, a fundamental problem in both information theory and many real-world applications. We begin by reviewing the relationship between variational autoencoders (VAEs), a powerful class of deep generative models, and rate-distortion theory, the theoretical foundation for lossy compression. By combining the ResNet VAE architecture and techniques including test-time quantization and quantization-aware training, we present a quantization-aware ResNet VAE (QARV) framework for lossy image compression. For sake of practical usage, we propose a new neural network architecture for fast decoding, and we introduce an adaptive normalization operation for variable-rate coding. QARV employs a hierarchical progressive coding structure, supports continuously variable-rate compression with fast entropy coding, and gives a better rate-distortion efficiency than existing baseline methods. Code is made publicly available at https://github.com/duanzhiihao/lossy-vae
Conventional cameras capture image irradiance on a sensor and convert it to RGB images using an image signal processor (ISP). The images can then be used for photography or visual computing tasks in a variety of applications, such as public safety surveillance and autonomous driving. One can argue that since RAW images contain all the captured information, the conversion of RAW to RGB using an ISP is not necessary for visual computing. In this paper, we propose a novel $\rho$-Vision framework to perform high-level semantic understanding and low-level compression using RAW images without the ISP subsystem used for decades. Considering the scarcity of available RAW image datasets, we first develop an unpaired CycleR2R network based on unsupervised CycleGAN to train modular unrolled ISP and inverse ISP (invISP) models using unpaired RAW and RGB images. We can then flexibly generate simulated RAW images (simRAW) using any existing RGB image dataset and finetune different models originally trained for the RGB domain to process real-world camera RAW images. We demonstrate object detection and image compression capabilities in RAW-domain using RAW-domain YOLOv3 and RAW image compressor (RIC) on snapshots from various cameras. Quantitative results reveal that RAW-domain task inference provides better detection accuracy and compression compared to RGB-domain processing. Furthermore, the proposed \r{ho}-Vision generalizes across various camera sensors and different task-specific models. Additional advantages of the proposed $\rho$-Vision that eliminates the ISP are the potential reductions in computations and processing times.
Bird's-Eye-View (BEV) 3D Object Detection is a crucial multi-view technique for autonomous driving systems. Recently, plenty of works are proposed, following a similar paradigm consisting of three essential components, i.e., camera feature extraction, BEV feature construction, and task heads. Among the three components, BEV feature construction is BEV-specific compared with 2D tasks. Existing methods aggregate the multi-view camera features to the flattened grid in order to construct the BEV feature. However, flattening the BEV space along the height dimension fails to emphasize the informative features of different heights. For example, the barrier is located at a low height while the truck is located at a high height. In this paper, we propose a novel method named BEV Slice Attention Network (BEV-SAN) for exploiting the intrinsic characteristics of different heights. Instead of flattening the BEV space, we first sample along the height dimension to build the global and local BEV slices. Then, the features of BEV slices are aggregated from the camera features and merged by the attention mechanism. Finally, we fuse the merged local and global BEV features by a transformer to generate the final feature map for task heads. The purpose of local BEV slices is to emphasize informative heights. In order to find them, we further propose a LiDAR-guided sampling strategy to leverage the statistical distribution of LiDAR to determine the heights of local slices. Compared with uniform sampling, LiDAR-guided sampling can determine more informative heights. We conduct detailed experiments to demonstrate the effectiveness of BEV-SAN. Code will be released.
Recently, Bird's-Eye-View (BEV) representation has gained increasing attention in multi-view 3D object detection, which has demonstrated promising applications in autonomous driving. Although multi-view camera systems can be deployed at low cost, the lack of depth information makes current approaches adopt large models for good performance. Therefore, it is essential to improve the efficiency of BEV 3D object detection. Knowledge Distillation (KD) is one of the most practical techniques to train efficient yet accurate models. However, BEV KD is still under-explored to the best of our knowledge. Different from image classification tasks, BEV 3D object detection approaches are more complicated and consist of several components. In this paper, we propose a unified framework named BEV-LGKD to transfer the knowledge in the teacher-student manner. However, directly applying the teacher-student paradigm to BEV features fails to achieve satisfying results due to heavy background information in RGB cameras. To solve this problem, we propose to leverage the localization advantage of LiDAR points. Specifically, we transform the LiDAR points to BEV space and generate the foreground mask and view-dependent mask for the teacher-student paradigm. It is to be noted that our method only uses LiDAR points to guide the KD between RGB models. As the quality of depth estimation is crucial for BEV perception, we further introduce depth distillation to our framework. Our unified framework is simple yet effective and achieves a significant performance boost. Code will be released.
Vision-Centric Bird-Eye-View (BEV) perception has shown promising potential and attracted increasing attention in autonomous driving. Recent works mainly focus on improving efficiency or accuracy but neglect the domain shift problem, resulting in severe degradation of transfer performance. With extensive observations, we figure out the significant domain gaps existing in the scene, weather, and day-night changing scenarios and make the first attempt to solve the domain adaption problem for multi-view 3D object detection. Since BEV perception approaches are usually complicated and contain several components, the domain shift accumulation on multi-latent spaces makes BEV domain adaptation challenging. In this paper, we propose a novel Multi-level Multi-space Alignment Teacher-Student ($M^{2}ATS$) framework to ease the domain shift accumulation, which consists of a Depth-Aware Teacher (DAT) and a Multi-space Feature Aligned (MFA) student model. Specifically, DAT model adopts uncertainty guidance to sample reliable depth information in target domain. After constructing domain-invariant BEV perception, it then transfers pixel and instance-level knowledge to student model. To further alleviate the domain shift at the global level, MFA student model is introduced to align task-relevant multi-space features of two domains. To verify the effectiveness of $M^{2}ATS$, we conduct BEV 3D object detection experiments on four cross domain scenarios and achieve state-of-the-art performance (e.g., +12.6% NDS and +9.1% mAP on Day-Night). Code and dataset will be released.
Quantizing floating-point neural network to its fixed-point representation is crucial for Learned Image Compression (LIC) because it ensures the decoding consistency for interoperability and reduces space-time complexity for implementation. Existing solutions often have to retrain the network for model quantization which is time consuming and impractical. This work suggests the use of Post-Training Quantization (PTQ) to directly process pretrained, off-the-shelf LIC models. We theoretically prove that minimizing the mean squared error (MSE) in PTQ is sub-optimal for compression task and thus develop a novel Rate-Distortion (R-D) Optimized PTQ (RDO-PTQ) to best retain the compression performance. Such RDO-PTQ just needs to compress few images (e.g., 10) to optimize the transformation of weight, bias, and activation of underlying LIC model from its native 32-bit floating-point (FP32) format to 8-bit fixed-point (INT8) precision for fixed-point inference onwards. Experiments reveal outstanding efficiency of the proposed method on different LICs, showing the closest coding performance to their floating-point counterparts. And, our method is a lightweight and plug-and-play approach without any need of model retraining which is attractive to practitioners.
A network based on complementary consistency training (CC-Net) is proposed for semi-supervised left atrial image segmentation in this paper. From the perspective of complementary information, CC-Net effectively utilizes unlabeled data and resolves the problem that semi-supervised segmentation algorithms currently in use have a limited capacity to extract information from unlabeled data. A primary model and two complementary auxiliary models are part of the complementary symmetric structure of the CC-Net. A complementary consistency training is formed by the inter-model perturbation between the primary model and the auxiliary models. The main model is better able to concentrate on the ambiguous region due to the complementary information provided by the two auxiliary models. Additionally, forcing consistency between the primary model and the auxiliary models makes it easier to obtain decision boundaries with little uncertainty. CC-Net was validated in the benchmark dataset of 2018 left atrial segmentation challenge, reaching Dice of 89.42% with 10% labeled data training and 91.14% with 20% labeled data training. By comparing with current state-of-the-art algorithms, CC-Net has the best segmentation performance and robustness. Our code is publicly available at https://github.com/Cuthbert-Huang/CC-Net.
Depth estimation is essential for various important real-world applications such as autonomous driving. However, it suffers from severe performance degradation in high-velocity scenario since traditional cameras can only capture blurred images. To deal with this problem, the spike camera is designed to capture the pixel-wise luminance intensity at high frame rate. However, depth estimation with spike camera remains very challenging using traditional monocular or stereo depth estimation algorithms, which are based on the photometric consistency. In this paper, we propose a novel Uncertainty-Guided Depth Fusion (UGDF) framework to fuse the predictions of monocular and stereo depth estimation networks for spike camera. Our framework is motivated by the fact that stereo spike depth estimation achieves better results at close range while monocular spike depth estimation obtains better results at long range. Therefore, we introduce a dual-task depth estimation architecture with a joint training strategy and estimate the distributed uncertainty to fuse the monocular and stereo results. In order to demonstrate the advantage of spike depth estimation over traditional camera depth estimation, we contribute a spike-depth dataset named CitySpike20K, which contains 20K paired samples, for spike depth estimation. UGDF achieves state-of-the-art results on CitySpike20K, surpassing all monocular or stereo spike depth estimation baselines. We conduct extensive experiments to evaluate the effectiveness and generalization of our method on CitySpike20K. To the best of our knowledge, our framework is the first dual-task fusion framework for spike camera depth estimation. Code and dataset will be released.
Recent work has shown a strong theoretical connection between variational autoencoders (VAEs) and the rate distortion theory. Motivated by this, we consider the problem of lossy image compression from the perspective of generative modeling. Starting from ResNet VAEs, which are originally designed for data (image) distribution modeling, we redesign their latent variable model using a quantization-aware posterior and prior, enabling easy quantization and entropy coding for image compression. Along with improved neural network blocks, we present a powerful and efficient class of lossy image coders, outperforming previous methods on natural image (lossy) compression. Our model compresses images in a coarse-to-fine fashion and supports parallel encoding and decoding, leading to fast execution on GPUs.