Neil M. Robertson

Xinshao Wang, Yang Hua, Elyor Kodirov et al.

There is a family of label modification approaches including self and non-self label correction (LC), and output regularisation. They are widely used for training robust deep neural networks (DNNs), but have not been mathematically and thoroughly analysed together. We study them and discover three key issues: (1) We are more interested in adopting Self LC as it leverages its own knowledge and requires no auxiliary models. However, it is unclear how to adaptively trust a learner as the training proceeds. (2) Some methods penalise while the others reward low-entropy (i.e., high-confidence) predictions, prompting us to ask which one is better. (3) Using the standard training setting, a learned model becomes less confident when severe noise exists. Self LC using high-entropy knowledge would generate high-entropy targets. To resolve the issue (1), inspired by a well-accepted finding, i.e., deep neural networks learn meaningful patterns before fitting noise, we propose a novel end-to-end method named ProSelfLC, which is designed according to the learning time and prediction entropy. Concretely, for any data point, we progressively and adaptively trust its predicted probability distribution versus its annotated one if a network has been trained for a relatively long time and the prediction is of low entropy. For the issue (2), the effectiveness of ProSelfLC defends entropy minimisation. By ProSelfLC, we empirically prove that it is more effective to redefine a semantic low-entropy state and optimise the learner toward it. To address the issue (3), we decrease the entropy of self knowledge using a low temperature before exploiting it to correct labels, so that the revised labels redefine low-entropy target probability distributions. We demonstrate the effectiveness of ProSelfLC through extensive experiments in both clean and noisy settings, and on both image and protein datasets.

Guanxiong Sun, Yang Hua, Guosheng Hu et al.

Recently, one-stage detectors have achieved competitive accuracy and faster speed compared with traditional two-stage detectors on image data. However, in the field of video object detection (VOD), most existing VOD methods are still based on two-stage detectors. Moreover, directly adapting existing VOD methods to one-stage detectors introduces unaffordable computational costs. In this paper, we first analyse the computational bottlenecks of using one-stage detectors for VOD. Based on the analysis, we present a simple yet efficient framework to address the computational bottlenecks and achieve efficient one-stage VOD by exploiting the temporal consistency in video frames. Specifically, our method consists of a location-prior network to filter out background regions and a size-prior network to skip unnecessary computations on low-level feature maps for specific frames. We test our method on various modern one-stage detectors and conduct extensive experiments on the ImageNet VID dataset. Excellent experimental results demonstrate the superior effectiveness, efficiency, and compatibility of our method. The code is available at https://github.com/guanxiongsun/vfe.pytorch.

Guanxiong Sun, Yang Hua, Guosheng Hu et al.

Deep video models, for example, 3D CNNs or video transformers, have achieved promising performance on sparse video tasks, i.e., predicting one result per video. However, challenges arise when adapting existing deep video models to dense video tasks, i.e., predicting one result per frame. Specifically, these models are expensive for deployment, less effective when handling redundant frames, and difficult to capture long-range temporal correlations. To overcome these issues, we propose a Temporal Dilated Video Transformer (TDViT) that consists of carefully designed temporal dilated transformer blocks (TDTB). TDTB can efficiently extract spatiotemporal representations and effectively alleviate the negative effect of temporal redundancy. Furthermore, by using hierarchical TDTBs, our approach obtains an exponentially expanded temporal receptive field and therefore can model long-range dynamics. Extensive experiments are conducted on two different dense video benchmarks, i.e., ImageNet VID for video object detection and YouTube VIS for video instance segmentation. Excellent experimental results demonstrate the superior efficiency, effectiveness, and compatibility of our method. The code is available at https://github.com/guanxiongsun/vfe.pytorch.

Guanxiong Sun, Yang Hua, Guosheng Hu et al.

State-of-the-art video object detection methods maintain a memory structure, either a sliding window or a memory queue, to enhance the current frame using attention mechanisms. However, we argue that these memory structures are not efficient or sufficient because of two implied operations: (1) concatenating all features in memory for enhancement, leading to a heavy computational cost; (2) frame-wise memory updating, preventing the memory from capturing more temporal information. In this paper, we propose a multi-level aggregation architecture via memory bank called MAMBA. Specifically, our memory bank employs two novel operations to eliminate the disadvantages of existing methods: (1) light-weight key-set construction which can significantly reduce the computational cost; (2) fine-grained feature-wise updating strategy which enables our method to utilize knowledge from the whole video. To better enhance features from complementary levels, i.e., feature maps and proposals, we further propose a generalized enhancement operation (GEO) to aggregate multi-level features in a unified manner. We conduct extensive evaluations on the challenging ImageNetVID dataset. Compared with existing state-of-the-art methods, our method achieves superior performance in terms of both speed and accuracy. More remarkably, MAMBA achieves mAP of 83.7/84.6% at 12.6/9.1 FPS with ResNet-101. Code is available at https://github.com/guanxiongsun/vfe.pytorch.

Xinshao Wang, Yang Hua, Elyor Kodirov et al.

To train robust deep neural networks (DNNs), we systematically study several target modification approaches, which include output regularisation, self and non-self label correction (LC). Two key issues are discovered: (1) Self LC is the most appealing as it exploits its own knowledge and requires no extra models. However, how to automatically decide the trust degree of a learner as training goes is not well answered in the literature? (2) Some methods penalise while the others reward low-entropy predictions, prompting us to ask which one is better? To resolve the first issue, taking two well-accepted propositions--deep neural networks learn meaningful patterns before fitting noise [3] and minimum entropy regularisation principle [10]--we propose a novel end-to-end method named ProSelfLC, which is designed according to learning time and entropy. Specifically, given a data point, we progressively increase trust in its predicted label distribution versus its annotated one if a model has been trained for enough time and the prediction is of low entropy (high confidence). For the second issue, according to ProSelfLC, we empirically prove that it is better to redefine a meaningful low-entropy status and optimise the learner toward it. This serves as a defence of entropy minimisation. We demonstrate the effectiveness of ProSelfLC through extensive experiments in both clean and noisy settings. The source code is available at https://github.com/XinshaoAmosWang/ProSelfLC-CVPR2021. Keywords: entropy minimisation, maximum entropy, confidence penalty, self knowledge distillation, label correction, label noise, semi-supervised learning, output regularisation

Yonggang Li, Guosheng Hu, Yongtao Wang et al.

Data augmentation (DA) techniques aim to increase data variability, and thus train deep networks with better generalisation. The pioneering AutoAugment automated the search for optimal DA policies with reinforcement learning. However, AutoAugment is extremely computationally expensive, limiting its wide applicability. Followup works such as Population Based Augmentation (PBA) and Fast AutoAugment improved efficiency, but their optimization speed remains a bottleneck. In this paper, we propose Differentiable Automatic Data Augmentation (DADA) which dramatically reduces the cost. DADA relaxes the discrete DA policy selection to a differentiable optimization problem via Gumbel-Softmax. In addition, we introduce an unbiased gradient estimator, RELAX, leading to an efficient and effective one-pass optimization strategy to learn an efficient and accurate DA policy. We conduct extensive experiments on CIFAR-10, CIFAR-100, SVHN, and ImageNet datasets. Furthermore, we demonstrate the value of Auto DA in pre-training for downstream detection problems. Results show our DADA is at least one order of magnitude faster than the state-of-the-art while achieving very comparable accuracy. The code is available at https://github.com/VDIGPKU/DADA.

Hao Zhu, Yang Yuan, Guosheng Hu et al.

Deep embedding learning is expected to learn a metric space in which features have smaller maximal intra-class distance than minimal inter-class distance. In recent years, one research focus is to solve the open-set problem by discriminative deep embedding learning in the field of face recognition (FR) and person re-identification (re-ID). Apart from open-set problem, we find that imbalanced training data is another main factor causing the performance degradation of FR and re-ID, and data imbalance widely exists in the real applications. However, very little research explores why and how data imbalance influences the performance of FR and re-ID with softmax or its variants. In this work, we deeply investigate data imbalance in the perspective of neural network optimisation and feature distribution about softmax. We find one main reason of performance degradation caused by data imbalance is that the weights (from the penultimate fully-connected layer) are far from their class centers in feature space. Based on this investigation, we propose a unified framework, Imbalance-Robust Softmax (IR-Softmax), which can simultaneously solve the open-set problem and reduce the influence of data imbalance. IR-Softmax can generalise to any softmax and its variants (which are discriminative for open-set problem) by directly setting the weights as their class centers, naturally solving the data imbalance problem. In this work, we explicitly re-formulate two discriminative softmax (A-Softmax and AM-Softmax) under the framework of IR-Softmax. We conduct extensive experiments on FR databases (LFW, MegaFace) and re-ID database (Market-1501, Duke), and IR-Softmax outperforms many state-of-the-art methods.

Slimane Larabi, Neil M. Robertson

In this paper we deal with contour detection based on the recent image analogy principle which has been successfully used for super-resolution, texture and curves synthesis and interactive editing. Hand-drawn outlines are initially as benchmarks. Given such a reference image, we present a new method based on this expertise to locate contours of a query image in the same way that it is done for the reference (i.e by analogy). Applying a image analogies for contour detection using hand drawn images as leaning images cannot gives good result for any query image. The contour detection may be improved if we increase the number of learning images such that there will be exist similarity between query image and some reference images. In addition of the hardness of contours drawing task, this will increase considerably the time computation. We investigated in this work, how can we avoid this constraint in order to guaranty that all contour pixels will be located for any query image. Fourteen derived stereo patches, derived from a mathematical study, are the knowledge used in order to locate contours at different scales independently of the light conditions. Comprehensive experiments are conducted on different data sets (BSD 500, Horses of Weizmann). The obtained results show superior performance via precision and recall vs. hand-drawn contours at multiple resolutions to the reported state of the art.

Shiyang Yan, Yang Hua, Neil M. Robertson

Recently, several approaches have been proposed to solve language generation problems. Transformer is currently state-of-the-art seq-to-seq model in language generation. Reinforcement Learning (RL) is useful in solving exposure bias and the optimisation on non-differentiable metrics in seq-to-seq language learning. However, Transformer is hard to combine with RL as the costly computing resource is required for sampling. We tackle this problem by proposing an off-policy RL learning algorithm where a behaviour policy represented by GRUs performs the sampling. We reduce the high variance of importance sampling (IS) by applying the truncated relative importance sampling (TRIS) technique and Kullback-Leibler (KL)-control concept. TRIS is a simple yet effective technique, and there is a theoretical proof that KL-control helps to reduce the variance of IS. We formulate this off-policy RL based on self-critical sequence training. Specifically, we use a Transformer-based captioning model as the target policy and use an image-guided language auto-encoder as the behaviour policy to explore the environment. The proposed algorithm achieves state-of-the-art performance on the visual paragraph generation and improved results on image captioning.

Shiyang Yan, Yang Hua, Neil Robertson

Image description generation plays an important role in many real-world applications, such as image retrieval, automatic navigation, and disabled people support. A well-developed task of image description generation is image captioning, which usually generates a short captioning sentence and thus neglects many of fine-grained properties, e.g., the information of subtle objects and their relationships. In this paper, we study the visual paragraph generation, which can describe the image with a long paragraph containing rich details. Previous research often generates the paragraph via a hierarchical Recurrent Neural Network (RNN)-like model, which has complex memorising, forgetting and coupling mechanism. Instead, we propose a novel pure CNN model, ParaCNN, to generate visual paragraph using hierarchical CNN architecture with contextual information between sentences within one paragraph. The ParaCNN can generate an arbitrary length of a paragraph, which is more applicable in many real-world applications. Furthermore, to enable the ParaCNN to model paragraph comprehensively, we also propose an adversarial twin net training scheme. During training, we force the forwarding network's hidden features to be close to that of the backwards network by using adversarial training. During testing, we only use the forwarding network, which already includes the knowledge of the backwards network, to generate a paragraph. We conduct extensive experiments on the Stanford Visual Paragraph dataset and achieve state-of-the-art performance.

Xinshao Wang, Elyor Kodirov, Yang Hua et al.

Loss functions play a crucial role in deep metric learning thus a variety of them have been proposed. Some supervise the learning process by pairwise or tripletwise similarity constraints while others take advantage of structured similarity information among multiple data points. In this work, we approach deep metric learning from a novel perspective. We propose instance cross entropy (ICE) which measures the difference between an estimated instance-level matching distribution and its ground-truth one. ICE has three main appealing properties. Firstly, similar to categorical cross entropy (CCE), ICE has clear probabilistic interpretation and exploits structured semantic similarity information for learning supervision. Secondly, ICE is scalable to infinite training data as it learns on mini-batches iteratively and is independent of the training set size. Thirdly, motivated by our relative weight analysis, seamless sample reweighting is incorporated. It rescales samples' gradients to control the differentiation degree over training examples instead of truncating them by sample mining. In addition to its simplicity and intuitiveness, extensive experiments on three real-world benchmarks demonstrate the superiority of ICE.

Xinshao Wang, Elyor Kodirov, Yang Hua et al.

Set-based person re-identification (SReID) is a matching problem that aims to verify whether two sets are of the same identity (ID). Existing SReID models typically generate a feature representation per image and aggregate them to represent the set as a single embedding. However, they can easily be perturbed by noises--perceptually/semantically low quality images--which are inevitable due to imperfect tracking/detection systems, or overfit to trivial images. In this work, we present a novel and simple solution to this problem based on ID-aware quality that measures the perceptual and semantic quality of images guided by their ID information. Specifically, we propose an ID-aware Embedding that consists of two key components: (1) Feature learning attention that aims to learn robust image embeddings by focusing on 'medium' hard images. This way it can prevent overfitting to trivial images, and alleviate the influence of outliers. (2) Feature fusion attention is to fuse image embeddings in the set to obtain the set-level embedding. It ignores noisy information and pays more attention to discriminative images to aggregate more discriminative information. Experimental results on four datasets show that our method outperforms state-of-the-art approaches despite the simplicity of our approach.

Xinshao Wang, Elyor Kodirov, Yang Hua et al.

Real-world large-scale datasets usually contain noisy labels and are imbalanced. Therefore, we propose derivative manipulation (DM), a novel and general example weighting approach for training robust deep models under these adverse conditions. DM has two main merits. First, loss function and example weighting are common techniques in the literature. DM reveals their connection (a loss function does example weighting) and is a replacement of both. Second, despite that a loss defines an example weighting scheme by its derivative, in the loss design, we need to consider whether it is differentiable. Instead, DM is more flexible by directly modifying the derivative so that a loss can be a non-elementary format too. Technically, DM defines an emphasis density function by a derivative magnitude function. DM is generic in that diverse weighting schemes can be derived. Extensive experiments on both vision and language tasks prove DM's effectiveness.

Xinshao Wang, Yang Hua, Elyor Kodirov et al.

In this work, we study robust deep learning against abnormal training data from the perspective of example weighting built in empirical loss functions, i.e., gradient magnitude with respect to logits, an angle that is not thoroughly studied so far. Consequently, we have two key findings: (1) Mean Absolute Error (MAE) Does Not Treat Examples Equally. We present new observations and insightful analysis about MAE, which is theoretically proved to be noise-robust. First, we reveal its underfitting problem in practice. Second, we analyse that MAE's noise-robustness is from emphasising on uncertain examples instead of treating training samples equally, as claimed in prior work. (2) The Variance of Gradient Magnitude Matters. We propose an effective and simple solution to enhance MAE's fitting ability while preserving its noise-robustness. Without changing MAE's overall weighting scheme, i.e., what examples get higher weights, we simply change its weighting variance non-linearly so that the impact ratio between two examples are adjusted. Our solution is termed Improved MAE (IMAE). We prove IMAE's effectiveness using extensive experiments: image classification under clean labels, synthetic label noise, and real-world unknown noise.

Alessandro Borgia, Yang Hua, Elyor Kodirov et al.

Video-based person re-identification deals with the inherent difficulty of matching unregulated sequences with different length and with incomplete target pose/viewpoint structure. Common approaches operate either by reducing the problem to the still images case, facing a significant information loss, or by exploiting inter-sequence temporal dependencies as in Siamese Recurrent Neural Networks or in gait analysis. However, in all cases, the inter-sequences pose/viewpoint misalignment is not considered, and the existing spatial approaches are mostly limited to the still images context. To this end, we propose a novel approach that can exploit more effectively the rich video information, by accounting for the role that the changing pose/viewpoint factor plays in the sequences matching process. Specifically, our approach consists of two components. The first one attempts to complement the original pose-incomplete information carried by the sequences with synthetic GAN-generated images, and fuse their feature vectors into a more discriminative viewpoint-insensitive embedding, namely Weighted Fusion (WF). Another one performs an explicit pose-based alignment of sequence pairs to promote coherent feature matching, namely Weighted-Pose Regulation (WPR). Extensive experiments on two large video-based benchmark datasets show that our approach outperforms considerably existing methods.

Zhiwei Liu, Xiangyu Zhu, Guosheng Hu et al.

Recently, deep learning based facial landmark detection has achieved great success. Despite this, we notice that the semantic ambiguity greatly degrades the detection performance. Specifically, the semantic ambiguity means that some landmarks (e.g. those evenly distributed along the face contour) do not have clear and accurate definition, causing inconsistent annotations by annotators. Accordingly, these inconsistent annotations, which are usually provided by public databases, commonly work as the ground-truth to supervise network training, leading to the degraded accuracy. To our knowledge, little research has investigated this problem. In this paper, we propose a novel probabilistic model which introduces a latent variable, i.e. the 'real' ground-truth which is semantically consistent, to optimize. This framework couples two parts (1) training landmark detection CNN and (2) searching the 'real' ground-truth. These two parts are alternatively optimized: the searched 'real' ground-truth supervises the CNN training; and the trained CNN assists the searching of 'real' ground-truth. In addition, to recover the unconfidently predicted landmarks due to occlusion and low quality, we propose a global heatmap correction unit (GHCU) to correct outliers by considering the global face shape as a constraint. Extensive experiments on both image-based (300W and AFLW) and video-based (300-VW) databases demonstrate that our method effectively improves the landmark detection accuracy and achieves the state of the art performance.

Xinshao Wang, Yang Hua, Elyor Kodirov et al.

The objective of deep metric learning (DML) is to learn embeddings that can capture semantic similarity and dissimilarity information among data points. Existing pairwise or tripletwise loss functions used in DML are known to suffer from slow convergence due to a large proportion of trivial pairs or triplets as the model improves. To improve this, ranking-motivated structured losses are proposed recently to incorporate multiple examples and exploit the structured information among them. They converge faster and achieve state-of-the-art performance. In this work, we unveil two limitations of existing ranking-motivated structured losses and propose a novel ranked list loss to solve both of them. First, given a query, only a fraction of data points is incorporated to build the similarity structure. Consequently, some useful examples are ignored and the structure is less informative. To address this, we propose to build a set-based similarity structure by exploiting all instances in the gallery. The learning setting can be interpreted as few-shot retrieval: given a mini-batch, every example is iteratively used as a query, and the rest ones compose the gallery to search, i.e., the support set in few-shot setting. The rest examples are split into a positive set and a negative set. For every mini-batch, the learning objective of ranked list loss is to make the query closer to the positive set than to the negative set by a margin. Second, previous methods aim to pull positive pairs as close as possible in the embedding space. As a result, the intraclass data distribution tends to be extremely compressed. In contrast, we propose to learn a hypersphere for each class in order to preserve useful similarity structure inside it, which functions as regularisation. Extensive experiments demonstrate the superiority of our proposal by comparing with the state-of-the-art methods.

Soumya Shubhra Ghosh, Yang Hua, Sankha Subhra Mukherjee et al.

Despite the breakthroughs in quality of image enhancement, an end-to-end solution for simultaneous recovery of the finer texture details and sharpness for degraded images with low resolution is still unsolved. Some existing approaches focus on minimizing the pixel-wise reconstruction error which results in a high peak signal-to-noise ratio. The enhanced images fail to provide high-frequency details and are perceptually unsatisfying, i.e., they fail to match the quality expected in a photo-realistic image. In this paper, we present Image Enhancement Generative Adversarial Network (IEGAN), a versatile framework capable of inferring photo-realistic natural images for both artifact removal and super-resolution simultaneously. Moreover, we propose a new loss function consisting of a combination of reconstruction loss, feature loss and an edge loss counterpart. The feature loss helps to push the output image to the natural image manifold and the edge loss preserves the sharpness of the output image. The reconstruction loss provides low-level semantic information to the generator regarding the quality of the generated images compared to the original. Our approach has been experimentally proven to recover photo-realistic textures from heavily compressed low-resolution images on public benchmarks and our proposed high-resolution World100 dataset.

Xinshao Wang, Yang Hua, Elyor Kodirov et al.

Deep metric learning aims to learn a deep embedding that can capture the semantic similarity of data points. Given the availability of massive training samples, deep metric learning is known to suffer from slow convergence due to a large fraction of trivial samples. Therefore, most existing methods generally resort to sample mining strategies for selecting nontrivial samples to accelerate convergence and improve performance. In this work, we identify two critical limitations of the sample mining methods, and provide solutions for both of them. First, previous mining methods assign one binary score to each sample, i.e., dropping or keeping it, so they only selects a subset of relevant samples in a mini-batch. Therefore, we propose a novel sample mining method, called Online Soft Mining (OSM), which assigns one continuous score to each sample to make use of all samples in the mini-batch. OSM learns extended manifolds that preserve useful intraclass variances by focusing on more similar positives. Second, the existing methods are easily influenced by outliers as they are generally included in the mined subset. To address this, we introduce Class-Aware Attention (CAA) that assigns little attention to abnormal data samples. Furthermore, by combining OSM and CAA, we propose a novel weighted contrastive loss to learn discriminative embeddings. Extensive experiments on two fine-grained visual categorisation datasets and two video-based person re-identification benchmarks show that our method significantly outperforms the state-of-the-art.

Marcel Sheeny de Moraes, Sankha Mukherjee, Neil M Robertson

Human interaction recognition is a challenging problem in computer vision and has been researched over the years due to its important applications. With the development of deep models for the human pose estimation problem, this work aims to verify the effectiveness of using the human pose in order to recognize the human interaction in monocular videos. This paper developed a method based on 5 steps: detect each person in the scene, track them, retrieve the human pose, extract features based on the pose and finally recognize the interaction using a classifier. The Two-Person interaction dataset was used for the development of this methodology. Using a whole sequence evaluation approach it achieved 87.56% of average accuracy of all interaction. Yun, et at achieved 91.10% using the same dataset, however their methodology used the depth sensor to recognize the interaction. The methodology developed in this paper shows that an RGB camera can be as effective as depth cameras to recognize the interaction between two persons using the recent development of deep models to estimate the human pose.