Xiang Li

Xiaoqian Shen, Xiang Li, Mohamed Elhoseiny

Video generation remains a challenging task due to spatiotemporal complexity and the requirement of synthesizing diverse motions with temporal consistency. Previous works attempt to generate videos in arbitrary lengths either in an autoregressive manner or regarding time as a continuous signal. However, they struggle to synthesize detailed and diverse motions with temporal coherence and tend to generate repetitive scenes after a few time steps. In this work, we argue that a single time-agnostic latent vector of style-based generator is insufficient to model various and temporally-consistent motions. Hence, we introduce additional time-dependent motion styles to model diverse motion patterns. In addition, a Motion Style Attention modulation mechanism, dubbed as MoStAtt, is proposed to augment frames with vivid dynamics for each specific scale (i.e., layer), which assigns attention score for each motion style w.r.t deconvolution filter weights in the target synthesis layer and softly attends different motion styles for weight modulation. Experimental results show our model achieves state-of-the-art performance on four unconditional $256^2$ video synthesis benchmarks trained with only 3 frames per clip and produces better qualitative results with respect to dynamic motions. Code and videos have been made available at https://github.com/xiaoqian-shen/MoStGAN-V.

Zhiqiang Yan, Kun Wang, Xiang Li et al.

Unsupervised depth completion aims to recover dense depth from the sparse one without using the ground-truth annotation. Although depth measurement obtained from LiDAR is usually sparse, it contains valid and real distance information, i.e., scale-consistent absolute depth values. Meanwhile, scale-agnostic counterparts seek to estimate relative depth and have achieved impressive performance. To leverage both the inherent characteristics, we thus suggest to model scale-consistent depth upon unsupervised scale-agnostic frameworks. Specifically, we propose the decomposed scale-consistent learning (DSCL) strategy, which disintegrates the absolute depth into relative depth prediction and global scale estimation, contributing to individual learning benefits. But unfortunately, most existing unsupervised scale-agnostic frameworks heavily suffer from depth holes due to the extremely sparse depth input and weak supervised signal. To tackle this issue, we introduce the global depth guidance (GDG) module, which attentively propagates dense depth reference into the sparse target via novel dense-to-sparse attention. Extensive experiments show the superiority of our method on outdoor KITTI benchmark, ranking 1st and outperforming the best KBNet more than 12% in RMSE. In addition, our approach achieves state-of-the-art performance on indoor NYUv2 dataset.

Changfeng Feng, Zhenyuan Chen, Xiang Li et al.

Object detection from aerial platforms under adverse atmospheric conditions, particularly haze, is paramount for robust drone autonomy. Yet, this domain remains largely underexplored, primarily hindered by the absence of specialized benchmarks. To bridge this gap, we present \textit{HazyDet}, the first, large-scale benchmark specifically designed for drone-view object detection in hazy conditions. Comprising 383,000 real-world instances derived from both naturally hazy captures and synthetically hazed scenes augmented from clear images, HazyDet provides a challenging and realistic testbed for advancing detection algorithms. To address the severe visual degradation induced by haze, we propose the Depth-Conditioned Detector (DeCoDet), a novel architecture that integrates a Depth-Conditioned Kernel to dynamically modulate feature representations based on depth cues. The practical efficacy and robustness of DeCoDet are further enhanced by its training with a Progressive Domain Fine-Tuning (PDFT) strategy to navigate synthetic-to-real domain shifts, and a Scale-Invariant Refurbishment Loss (SIRLoss) to ensure resilient learning from potentially noisy depth annotations. Comprehensive empirical validation on HazyDet substantiates the superiority of our unified DeCoDet framework, which achieves state-of-the-art performance, surpassing the closest competitor by a notable +1.5\% mAP on challenging real-world hazy test scenarios. Our dataset and toolkit are available at https://github.com/GrokCV/HazyDet.

Jiayi Yuan, Haobo Jiang, Xiang Li et al.

Image guidance is an effective strategy for depth super-resolution. Generally, most existing methods employ hand-crafted operators to decompose the high-frequency (HF) and low-frequency (LF) ingredients from low-resolution depth maps and guide the HF ingredients by directly concatenating them with image features. However, the hand-designed operators usually cause inferior HF maps (e.g., distorted or structurally missing) due to the diverse appearance of complex depth maps. Moreover, the direct concatenation often results in weak guidance because not all image features have a positive effect on the HF maps. In this paper, we develop a recurrent structure attention guided (RSAG) framework, consisting of two important parts. First, we introduce a deep contrastive network with multi-scale filters for adaptive frequency-domain separation, which adopts contrastive networks from large filters to small ones to calculate the pixel contrasts for adaptive high-quality HF predictions. Second, instead of the coarse concatenation guidance, we propose a recurrent structure attention block, which iteratively utilizes the latest depth estimation and the image features to jointly select clear patterns and boundaries, aiming at providing refined guidance for accurate depth recovery. In addition, we fuse the features of HF maps to enhance the edge structures in the decomposed LF maps. Extensive experiments show that our approach obtains superior performance compared with state-of-the-art depth super-resolution methods.

Jiayi Yuan, Haobo Jiang, Xiang Li et al.

Real depth super-resolution (DSR), unlike synthetic settings, is a challenging task due to the structural distortion and the edge noise caused by the natural degradation in real-world low-resolution (LR) depth maps. These defeats result in significant structure inconsistency between the depth map and the RGB guidance, which potentially confuses the RGB-structure guidance and thereby degrades the DSR quality. In this paper, we propose a novel structure flow-guided DSR framework, where a cross-modality flow map is learned to guide the RGB-structure information transferring for precise depth upsampling. Specifically, our framework consists of a cross-modality flow-guided upsampling network (CFUNet) and a flow-enhanced pyramid edge attention network (PEANet). CFUNet contains a trilateral self-attention module combining both the geometric and semantic correlations for reliable cross-modality flow learning. Then, the learned flow maps are combined with the grid-sampling mechanism for coarse high-resolution (HR) depth prediction. PEANet targets at integrating the learned flow map as the edge attention into a pyramid network to hierarchically learn the edge-focused guidance feature for depth edge refinement. Extensive experiments on real and synthetic DSR datasets verify that our approach achieves excellent performance compared to state-of-the-art methods.

Ge Wu, Xin Zhang, Zheng Li et al.

Prompt learning has surfaced as an effective approach to enhance the performance of Vision-Language Models (VLMs) like CLIP when applied to downstream tasks. However, current learnable prompt tokens are primarily used for the single phase of adapting to tasks (i.e., adapting prompt), easily leading to overfitting risks. In this work, we propose a novel Cascade Prompt Learning CasPL framework to enable prompt learning to serve both generic and specific expertise (i.e., boosting and adapting prompt) simultaneously. Specifically, CasPL is a new learning paradigm comprising two distinct phases of learnable prompts: the first boosting prompt is crafted to extract domain-general knowledge from a senior larger CLIP teacher model by aligning their predicted logits using extensive unlabeled domain images. The second adapting prompt is then cascaded with the frozen first set to fine-tune the downstream tasks, following the approaches employed in prior research. In this manner, CasPL can effectively capture both domain-general and task-specific representations into explicitly different gradual groups of prompts, thus potentially alleviating overfitting issues in the target domain. It's worth noting that CasPL serves as a plug-and-play module that can seamlessly integrate into any existing prompt learning approach. CasPL achieves a significantly better balance between performance and inference speed, which is especially beneficial for deploying smaller VLM models in resource-constrained environments. Compared to the previous state-of-the-art method PromptSRC, CasPL shows an average improvement of 1.85% for base classes, 3.44% for novel classes, and 2.72% for the harmonic mean over 11 image classification datasets. Code is publicly available at: https://github.com/megvii-research/CasPL.

Xiang Li, Yinpeng Chen, Chung-Ching Lin et al.

This paper presents a novel approach to object completion, with the primary goal of reconstructing a complete object from its partially visible components. Our method, named MaskComp, delineates the completion process through iterative stages of generation and segmentation. In each iteration, the object mask is provided as an additional condition to boost image generation, and, in return, the generated images can lead to a more accurate mask by fusing the segmentation of images. We demonstrate that the combination of one generation and one segmentation stage effectively functions as a mask denoiser. Through alternation between the generation and segmentation stages, the partial object mask is progressively refined, providing precise shape guidance and yielding superior object completion results. Our experiments demonstrate the superiority of MaskComp over existing approaches, e.g., ControlNet and Stable Diffusion, establishing it as an effective solution for object completion.

Xinbin Yuan, Zhaohui Zheng, Yuxuan Li et al.

While witnessed with rapid development, remote sensing object detection remains challenging for detecting high aspect ratio objects. This paper shows that large strip convolutions are good feature representation learners for remote sensing object detection and can detect objects of various aspect ratios well. Based on large strip convolutions, we build a new network architecture called Strip R-CNN, which is simple, efficient, and powerful. Unlike recent remote sensing object detectors that leverage large-kernel convolutions with square shapes, our Strip R-CNN takes advantage of sequential orthogonal large strip convolutions in our backbone network StripNet to capture spatial information. In addition, we improve the localization capability of remote-sensing object detectors by decoupling the detection heads and equipping the localization branch with strip convolutions in our strip head. Extensive experiments on several benchmarks, for example DOTA, FAIR1M, HRSC2016, and DIOR, show that our Strip R-CNN can greatly improve previous work. In particular, our 30M model achieves 82.75% mAP on DOTA-v1.0, setting a new state-of-the-art record. Our code will be made publicly available.Code is available at https://github.com/YXB-NKU/Strip-R-CNN.

Zheng Li, Xiang Li, Lingfeng Yang et al.

Knowledge distillation usually transfers the knowledge from a pre-trained cumbersome teacher network to a compact student network, which follows the classical teacher-teaching-student paradigm. Based on this paradigm, previous methods mostly focus on how to efficiently train a better student network for deployment. Different from the existing practices, in this paper, we propose a novel student-helping-teacher formula, Teacher Evolution via Self-Knowledge Distillation (TESKD), where the target teacher (for deployment) is learned with the help of multiple hierarchical students by sharing the structural backbone. The diverse feedback from multiple students allows the teacher to improve itself through the shared feature representations. The effectiveness of our proposed framework is demonstrated by extensive experiments with various network settings on two standard benchmarks including CIFAR-100 and ImageNet. Notably, when trained together with our proposed method, ResNet-18 achieves 79.15% and 71.14% accuracy on CIFAR-100 and ImageNet, outperforming the baseline results by 4.74% and 1.43%, respectively. The code is available at: https://github.com/zhengli427/TESKD.

Wenhai Wang, Enze Xie, Xiang Li et al.

Transformer recently has presented encouraging progress in computer vision. In this work, we present new baselines by improving the original Pyramid Vision Transformer (PVT v1) by adding three designs, including (1) linear complexity attention layer, (2) overlapping patch embedding, and (3) convolutional feed-forward network. With these modifications, PVT v2 reduces the computational complexity of PVT v1 to linear and achieves significant improvements on fundamental vision tasks such as classification, detection, and segmentation. Notably, the proposed PVT v2 achieves comparable or better performances than recent works such as Swin Transformer. We hope this work will facilitate state-of-the-art Transformer researches in computer vision. Code is available at https://github.com/whai362/PVT.

Li Xiang, Chen Shuo, Xia Yan et al.

The merits of fast convergence and potentially better performance of the weight normalization family have drawn increasing attention in recent years. These methods use standardization or normalization that changes the weight $\boldsymbol{W}$ to $\boldsymbol{W}'$, which makes $\boldsymbol{W}'$ independent to the magnitude of $\boldsymbol{W}$. Surprisingly, $\boldsymbol{W}$ must be decayed during gradient descent, otherwise we will observe a severe under-fitting problem, which is very counter-intuitive since weight decay is widely known to prevent deep networks from over-fitting. In this paper, we \emph{theoretically} prove that the weight decay term $\frac{1}{2}λ||{\boldsymbol{W}}||^2$ merely modulates the effective learning rate for improving objective optimization, and has no influence on generalization when the weight normalization family is compositely employed. Furthermore, we also expose several critical problems when introducing weight decay term to weight normalization family, including the missing of global minimum and training instability. To address these problems, we propose an $ε-$shifted $L_2$ regularizer, which shifts the $L_2$ objective by a positive constant $ε$. Such a simple operation can theoretically guarantee the existence of global minimum, while preventing the network weights from being too small and thus avoiding gradient float overflow. It significantly improves the training stability and can achieve slightly better performance in our practice. The effectiveness of $ε-$shifted $L_2$ regularizer is comprehensively validated on the ImageNet, CIFAR-100, and COCO datasets. Our codes and pretrained models will be released in https://github.com/implus/PytorchInsight.

Xiang Li, Xiaolin Hu, Jian Yang

The Convolutional Neural Networks (CNNs) generate the feature representation of complex objects by collecting hierarchical and different parts of semantic sub-features. These sub-features can usually be distributed in grouped form in the feature vector of each layer, representing various semantic entities. However, the activation of these sub-features is often spatially affected by similar patterns and noisy backgrounds, resulting in erroneous localization and identification. We propose a Spatial Group-wise Enhance (SGE) module that can adjust the importance of each sub-feature by generating an attention factor for each spatial location in each semantic group, so that every individual group can autonomously enhance its learnt expression and suppress possible noise. The attention factors are only guided by the similarities between the global and local feature descriptors inside each group, thus the design of SGE module is extremely lightweight with \emph{almost no extra parameters and calculations}. Despite being trained with only category supervisions, the SGE component is extremely effective in highlighting multiple active areas with various high-order semantics (such as the dog's eyes, nose, etc.). When integrated with popular CNN backbones, SGE can significantly boost the performance of image recognition tasks. Specifically, based on ResNet50 backbones, SGE achieves 1.2\% Top-1 accuracy improvement on the ImageNet benchmark and 1.0$\sim$2.0\% AP gain on the COCO benchmark across a wide range of detectors (Faster/Mask/Cascade RCNN and RetinaNet). Codes and pretrained models are available at https://github.com/implus/PytorchInsight.

Yi Pan, Hanqi Jiang, Junhao Chen et al.

Neuromorphic computing has emerged as a promising energy-efficient alternative to traditional artificial intelligence, predominantly utilizing spiking neural networks (SNNs) implemented on neuromorphic hardware. Significant advancements have been made in SNN-based convolutional neural networks (CNNs) and Transformer architectures. However, neuromorphic computing for the medical imaging domain remains underexplored. In this study, we introduce EG-SpikeFormer, an SNN architecture tailored for clinical tasks that incorporates eye-gaze data to guide the model's attention to the diagnostically relevant regions in medical images. Our developed approach effectively addresses shortcut learning issues commonly observed in conventional models, especially in scenarios with limited clinical data and high demands for model reliability, generalizability, and transparency. Our EG-SpikeFormer not only demonstrates superior energy efficiency and performance in medical image prediction tasks but also enhances clinical relevance through multi-modal information alignment. By incorporating eye-gaze data, the model improves interpretability and generalization, opening new directions for applying neuromorphic computing in healthcare.

Chenxu Peng, Chenxu Wang, Minrui Zou et al.

Infrared object tracking plays a crucial role in Anti-Unmanned Aerial Vehicle (Anti-UAV) applications. Existing trackers often depend on cropped template regions and have limited motion modeling capabilities, which pose challenges when dealing with tiny targets. To address this, we propose a simple yet effective infrared tiny-object tracker that enhances tracking performance by integrating global detection and motion-aware learning with temporal priors. Our method is based on object detection and achieves significant improvements through two key innovations. First, we introduce frame dynamics, leveraging frame difference and optical flow to encode both prior target features and motion characteristics at the input level, enabling the model to better distinguish the target from background clutter. Second, we propose a trajectory constraint filtering strategy in the post-processing stage, utilizing spatio-temporal priors to suppress false positives and enhance tracking robustness. Extensive experiments show that our method consistently outperforms existing approaches across multiple metrics in challenging infrared UAV tracking scenarios. Notably, we achieve state-of-the-art performance in the 4th Anti-UAV Challenge, securing 1st place in Track 1 and 2nd place in Track 2.

Qiong Chang, Xiang Li, Xin Xu et al.

We present a lightweight system for stereo matching through embedded GPUs. It breaks the trade-off between accuracy and processing speed in stereo matching, enabling our embedded system to further improve the matching accuracy while ensuring real-time processing. The main idea of our method is to construct a tiny neural network based on variational auto-encoder (VAE) to upsample and refinement a small size of coarse disparity map, which is first generated by a traditional matching method. The proposed hybrid structure cannot only bring the advantage of traditional methods in terms of computational complexity, but also ensure the matching accuracy under the impact of neural network. Extensive experiments on the KITTI 2015 benchmark demonstrate that our tiny system exhibits high robustness in improving the accuracy of the coarse disparity maps generated by different algorithms, while also running in real-time on embedded GPUs.

Xiang Li, Congcong Wen, Yuan Hu et al.

The remarkable achievements of ChatGPT and GPT-4 have sparked a wave of interest and research in the field of large language models for Artificial General Intelligence (AGI). These models provide intelligent solutions close to human thinking, enabling us to use general artificial intelligence to solve problems in various applications. However, in remote sensing (RS), the scientific literature on the implementation of AGI remains relatively scant. Existing AI-related research in remote sensing primarily focuses on visual understanding tasks while neglecting the semantic understanding of the objects and their relationships. This is where vision-language models excel, as they enable reasoning about images and their associated textual descriptions, allowing for a deeper understanding of the underlying semantics. Vision-language models can go beyond visual recognition of RS images, model semantic relationships, and generate natural language descriptions of the image. This makes them better suited for tasks requiring visual and textual understanding, such as image captioning, and visual question answering. This paper provides a comprehensive review of the research on vision-language models in remote sensing, summarizing the latest progress, highlighting challenges, and identifying potential research opportunities.

Zhiqiang Yan, Kun Wang, Xiang Li et al.

Depth completion deals with the problem of recovering dense depth maps from sparse ones, where color images are often used to facilitate this task. Recent approaches mainly focus on image guided learning frameworks to predict dense depth. However, blurry guidance in the image and unclear structure in the depth still impede the performance of the image guided frameworks. To tackle these problems, we explore a repetitive design in our image guided network to gradually and sufficiently recover depth values. Specifically, the repetition is embodied in both the image guidance branch and depth generation branch. In the former branch, we design a repetitive hourglass network to extract discriminative image features of complex environments, which can provide powerful contextual instruction for depth prediction. In the latter branch, we introduce a repetitive guidance module based on dynamic convolution, in which an efficient convolution factorization is proposed to simultaneously reduce its complexity and progressively model high-frequency structures. Extensive experiments show that our method achieves superior or competitive results on KITTI benchmark and NYUv2 dataset.

Wenhai Wang, Enze Xie, Xiang Li et al.

Scene text detection and recognition have been well explored in the past few years. Despite the progress, efficient and accurate end-to-end spotting of arbitrarily-shaped text remains challenging. In this work, we propose an end-to-end text spotting framework, termed PAN++, which can efficiently detect and recognize text of arbitrary shapes in natural scenes. PAN++ is based on the kernel representation that reformulates a text line as a text kernel (central region) surrounded by peripheral pixels. By systematically comparing with existing scene text representations, we show that our kernel representation can not only describe arbitrarily-shaped text but also well distinguish adjacent text. Moreover, as a pixel-based representation, the kernel representation can be predicted by a single fully convolutional network, which is very friendly to real-time applications. Taking the advantages of the kernel representation, we design a series of components as follows: 1) a computationally efficient feature enhancement network composed of stacked Feature Pyramid Enhancement Modules (FPEMs); 2) a lightweight detection head cooperating with Pixel Aggregation (PA); and 3) an efficient attention-based recognition head with Masked RoI. Benefiting from the kernel representation and the tailored components, our method achieves high inference speed while maintaining competitive accuracy. Extensive experiments show the superiority of our method. For example, the proposed PAN++ achieves an end-to-end text spotting F-measure of 64.9 at 29.2 FPS on the Total-Text dataset, which significantly outperforms the previous best method. Code will be available at: https://git.io/PAN.

Meimei Shang, Fei Gao, Xiang Li et al.

In this paper, we propose a novel method to learn face sketch synthesis models by using unpaired data. Our main idea is bridging the photo domain $\mathcal{X}$ and the sketch domain $Y$ by using the line-drawing domain $\mathcal{Z}$. Specially, we map both photos and sketches to line-drawings by using a neural style transfer method, i.e. $F: \mathcal{X}/\mathcal{Y} \mapsto \mathcal{Z}$. Consequently, we obtain \textit{pseudo paired data} $(\mathcal{Z}, \mathcal{Y})$, and can learn the mapping $G:\mathcal{Z} \mapsto \mathcal{Y}$ in a supervised learning manner. In the inference stage, given a facial photo, we can first transfer it to a line-drawing and then to a sketch by $G \circ F$. Additionally, we propose a novel stroke loss for generating different types of strokes. Our method, termed sRender, accords well with human artists' rendering process. Experimental results demonstrate that sRender can generate multi-style sketches, and significantly outperforms existing unpaired image-to-image translation methods.

Xiang Li, Shusen Wang, Kun Chen et al.

We study distributed computing of the truncated singular value decomposition problem. We develop an algorithm that we call \texttt{LocalPower} for improving communication efficiency. Specifically, we uniformly partition the dataset among $m$ nodes and alternate between multiple (precisely $p$) local power iterations and one global aggregation. In the aggregation, we propose to weight each local eigenvector matrix with orthogonal Procrustes transformation (OPT). As a practical surrogate of OPT, sign-fixing, which uses a diagonal matrix with $\pm 1$ entries as weights, has better computation complexity and stability in experiments. We theoretically show that under certain assumptions \texttt{LocalPower} lowers the required number of communications by a factor of $p$ to reach a constant accuracy. We also show that the strategy of periodically decaying $p$ helps obtain high-precision solutions. We conduct experiments to demonstrate the effectiveness of \texttt{LocalPower}.

Yuan Hu, Yunpeng Chen, Xiang Li et al.

Features from multiple scales can greatly benefit the semantic edge detection task if they are well fused. However, the prevalent semantic edge detection methods apply a fixed weight fusion strategy where images with different semantics are forced to share the same weights, resulting in universal fusion weights for all images and locations regardless of their different semantics or local context. In this work, we propose a novel dynamic feature fusion strategy that assigns different fusion weights for different input images and locations adaptively. This is achieved by a proposed weight learner to infer proper fusion weights over multi-level features for each location of the feature map, conditioned on the specific input. In this way, the heterogeneity in contributions made by different locations of feature maps and input images can be better considered and thus help produce more accurate and sharper edge predictions. We show that our model with the novel dynamic feature fusion is superior to fixed weight fusion and also the naïve location-invariant weight fusion methods, via comprehensive experiments on benchmarks Cityscapes and SBD. In particular, our method outperforms all existing well established methods and achieves new state-of-the-art.

Shuo Chen, Chen Gong, Jian Yang et al.

In the past decades, intensive efforts have been put to design various loss functions and metric forms for metric learning problem. These improvements have shown promising results when the test data is similar to the training data. However, the trained models often fail to produce reliable distances on the ambiguous test pairs due to the distribution bias between training set and test set. To address this problem, the Adversarial Metric Learning (AML) is proposed in this paper, which automatically generates adversarial pairs to remedy the distribution bias and facilitate robust metric learning. Specifically, AML consists of two adversarial stages, i.e. confusion and distinguishment. In confusion stage, the ambiguous but critical adversarial data pairs are adaptively generated to mislead the learned metric. In distinguishment stage, a metric is exhaustively learned to try its best to distinguish both the adversarial pairs and the original training pairs. Thanks to the challenges posed by the confusion stage in such competing process, the AML model is able to grasp plentiful difficult knowledge that has not been contained by the original training pairs, so the discriminability of AML can be significantly improved. The entire model is formulated into optimization framework, of which the global convergence is theoretically proved. The experimental results on toy data and practical datasets clearly demonstrate the superiority of AML to the representative state-of-the-art metric learning methodologies.

Wenhai Wang, Xiang Li, Jian Yang et al.

Basing on the analysis by revealing the equivalence of modern networks, we find that both ResNet and DenseNet are essentially derived from the same "dense topology", yet they only differ in the form of connection -- addition (dubbed "inner link") vs. concatenation (dubbed "outer link"). However, both two forms of connections have the superiority and insufficiency. To combine their advantages and avoid certain limitations on representation learning, we present a highly efficient and modularized Mixed Link Network (MixNet) which is equipped with flexible inner link and outer link modules. Consequently, ResNet, DenseNet and Dual Path Network (DPN) can be regarded as a special case of MixNet, respectively. Furthermore, we demonstrate that MixNets can achieve superior efficiency in parameter over the state-of-the-art architectures on many competitive datasets like CIFAR-10/100, SVHN and ImageNet.

Xiang Li, Shuo Chen, Xiaolin Hu et al.

This paper first answers the question "why do the two most powerful techniques Dropout and Batch Normalization (BN) often lead to a worse performance when they are combined together?" in both theoretical and statistical aspects. Theoretically, we find that Dropout would shift the variance of a specific neural unit when we transfer the state of that network from train to test. However, BN would maintain its statistical variance, which is accumulated from the entire learning procedure, in the test phase. The inconsistency of that variance (we name this scheme as "variance shift") causes the unstable numerical behavior in inference that leads to more erroneous predictions finally, when applying Dropout before BN. Thorough experiments on DenseNet, ResNet, ResNeXt and Wide ResNet confirm our findings. According to the uncovered mechanism, we next explore several strategies that modifies Dropout and try to overcome the limitations of their combination by avoiding the variance shift risks.

Jifeng Wang, Xiang Li, Le Hui et al.

Understanding shadows from a single image spontaneously derives into two types of task in previous studies, containing shadow detection and shadow removal. In this paper, we present a multi-task perspective, which is not embraced by any existing work, to jointly learn both detection and removal in an end-to-end fashion that aims at enjoying the mutually improved benefits from each other. Our framework is based on a novel STacked Conditional Generative Adversarial Network (ST-CGAN), which is composed of two stacked CGANs, each with a generator and a discriminator. Specifically, a shadow image is fed into the first generator which produces a shadow detection mask. That shadow image, concatenated with its predicted mask, goes through the second generator in order to recover its shadow-free image consequently. In addition, the two corresponding discriminators are very likely to model higher level relationships and global scene characteristics for the detected shadow region and reconstruction via removing shadows, respectively. More importantly, for multi-task learning, our design of stacked paradigm provides a novel view which is notably different from the commonly used one as the multi-branch version. To fully evaluate the performance of our proposed framework, we construct the first large-scale benchmark with 1870 image triplets (shadow image, shadow mask image, and shadow-free image) under 135 scenes. Extensive experimental results consistently show the advantages of ST-CGAN over several representative state-of-the-art methods on two large-scale publicly available datasets and our newly released one.