The objective learning formulation is essential for the success of convolutional neural networks. In this work, we analyse thoroughly the standard learning objective functions for multi-class classification CNNs: softmax regression (SR) for single-label scenario and logistic regression (LR) for multi-label scenario. Our analyses lead to an inspiration of exploiting LR for single-label classification learning, and then the disclosing of the negative class distraction problem in LR. To address this problem, we develop two novel LR based objective functions that not only generalise the conventional LR but importantly turn out to be competitive alternatives to SR in single label classification. Extensive comparative evaluations demonstrate the model learning advantages of the proposed LR functions over the commonly adopted SR in single-label coarse-grained object categorisation and cross-class fine-grained person instance identification tasks. We also show the performance superiority of our method on clothing attribute classification in comparison to the vanilla LR function.
Knowledge distillation is an effective approach to transferring knowledge from a teacher neural network to a student target network for satisfying the low-memory and fast running requirements in practice use. Whilst being able to create stronger target networks compared to the vanilla non-teacher based learning strategy, this scheme needs to train additionally a large teacher model with expensive computational cost. In this work, we present a Self-Referenced Deep Learning (SRDL) strategy. Unlike both vanilla optimisation and existing knowledge distillation, SRDL distils the knowledge discovered by the in-training target model back to itself to regularise the subsequent learning procedure therefore eliminating the need for training a large teacher model. SRDL improves the model generalisation performance compared to vanilla learning and conventional knowledge distillation approaches with negligible extra computational cost. Extensive evaluations show that a variety of deep networks benefit from SRDL resulting in enhanced deployment performance on both coarse-grained object categorisation tasks (CIFAR10, CIFAR100, Tiny ImageNet, and ImageNet) and fine-grained person instance identification tasks (Market-1501).
Existing logo detection benchmarks consider artificial deployment scenarios by assuming that large training data with fine-grained bounding box annotations for each class are available for model training. Such assumptions are often invalid in realistic logo detection scenarios where new logo classes come progressively and require to be detected with little or none budget for exhaustively labelling fine-grained training data for every new class. Existing benchmarks are thus unable to evaluate the true performance of a logo detection method in realistic and open deployments. In this work, we introduce a more realistic and challenging logo detection setting, called Open Logo Detection. Specifically, this new setting assumes fine-grained labelling only on a small proportion of logo classes whilst the remaining classes have no labelled training data to simulate the open deployment. We further create an open logo detection benchmark, called OpenLogo,to promote the investigation of this new challenge. OpenLogo contains 27,083 images from 352 logo classes, built by aggregating/refining 7 existing datasets and establishing an open logo detection evaluation protocol. To address this challenge, we propose a Context Adversarial Learning (CAL) approach to synthesising training data with coherent logo instance appearance against diverse background context for enabling more effective optimisation of contemporary deep learning detection models. Experiments show the performance advantage of CAL over existing state-of-the-art alternative methods on the more realistic and challenging OpenLogo benchmark.
Existing vehicle re-identification (re-id) evaluation benchmarks consider strongly artificial test scenarios by assuming the availability of high quality images and fine-grained appearance at an almost constant image scale, reminiscent to images required for Automatic Number Plate Recognition, e.g. VeRi-776. Such assumptions are often invalid in realistic vehicle re-id scenarios where arbitrarily changing image resolutions (scales) are the norm. This makes the existing vehicle re-id benchmarks limited for testing the true performance of a re-id method. In this work, we introduce a more realistic and challenging vehicle re-id benchmark, called Vehicle Re-Identification in Context (VRIC). In contrast to existing datasets, VRIC is uniquely characterised by vehicle images subject to more realistic and unconstrained variations in resolution (scale), motion blur, illumination, occlusion, and viewpoint. It contains 60,430 images of 5,622 vehicle identities captured by 60 different cameras at heterogeneous road traffic scenes in both day-time and night-time.
Knowledge distillation is effective to train small and generalisable network models for meeting the low-memory and fast running requirements. Existing offline distillation methods rely on a strong pre-trained teacher, which enables favourable knowledge discovery and transfer but requires a complex two-phase training procedure. Online counterparts address this limitation at the price of lacking a highcapacity teacher. In this work, we present an On-the-fly Native Ensemble (ONE) strategy for one-stage online distillation. Specifically, ONE trains only a single multi-branch network while simultaneously establishing a strong teacher on-the- fly to enhance the learning of target network. Extensive evaluations show that ONE improves the generalisation performance a variety of deep neural networks more significantly than alternative methods on four image classification dataset: CIFAR10, CIFAR100, SVHN, and ImageNet, whilst having the computational efficiency advantages.
Mostexistingpersonre-identification(re-id)methods relyon supervised model learning on per-camera-pair manually labelled pairwise training data. This leads to poor scalability in practical re-id deployment due to the lack of exhaustive identity labelling of image positive and negative pairs for every camera pair. In this work, we address this problem by proposing an unsupervised re-id deep learning approach capable of incrementally discovering and exploiting the underlying re-id discriminative information from automatically generated person tracklet data from videos in an end-to-end model optimisation. We formulate a Tracklet Association Unsupervised Deep Learning (TAUDL) framework characterised by jointly learning per-camera (within-camera) tracklet association (labelling) and cross-camera tracklet correlation by maximising the discovery of most likely tracklet relationships across camera views. Extensive experiments demonstrate the superiority of the proposed TAUDL model over the state-of-the-art unsupervised and domain adaptation re- id methods using six person re-id benchmarking datasets.
Face recognition (FR) is one of the most extensively investigated problems in computer vision. Significant progress in FR has been made due to the recent introduction of the larger scale FR challenges, particularly with constrained social media web images, e.g. high-resolution photos of celebrity faces taken by professional photo-journalists. However, the more challenging FR in unconstrained and low-resolution surveillance images remains largely under-studied. To facilitate more studies on developing FR models that are effective and robust for low-resolution surveillance facial images, we introduce a new Surveillance Face Recognition Challenge, which we call the QMUL-SurvFace benchmark. This new benchmark is the largest and more importantly the only true surveillance FR benchmark to our best knowledge, where low-resolution images are not synthesised by artificial down-sampling of native high-resolution images. This challenge contains 463,507 face images of 15,573 distinct identities captured in real-world uncooperative surveillance scenes over wide space and time. As a consequence, it presents an extremely challenging FR benchmark. We benchmark the FR performance on this challenge using five representative deep learning face recognition models, in comparison to existing benchmarks. We show that the current state of the arts are still far from being satisfactory to tackle the under-investigated surveillance FR problem in practical forensic scenarios. Face recognition is generally more difficult in an open-set setting which is typical for surveillance scenarios, owing to a large number of non-target people (distractors) appearing open spaced scenes. This is evidently so that on the new Surveillance FR Challenge, the top-performing CentreFace deep learning FR model on the MegaFace benchmark can now only achieve 13.2% success rate (at Rank-20) at a 10% false alarm rate.
Deep learning methods have started to dominate the research progress of video-based person re-identification (re-id). However, existing methods mostly consider supervised learning, which requires exhaustive manual efforts for labelling cross-view pairwise data. Therefore, they severely lack scalability and practicality in real-world video surveillance applications. In this work, to address the video person re-id task, we formulate a novel Deep Association Learning (DAL) scheme, the first end-to-end deep learning method using none of the identity labels in model initialisation and training. DAL learns a deep re-id matching model by jointly optimising two margin-based association losses in an end-to-end manner, which effectively constrains the association of each frame to the best-matched intra-camera representation and cross-camera representation. Existing standard CNNs can be readily employed within our DAL scheme. Experiment results demonstrate that our proposed DAL significantly outperforms current state-of-the-art unsupervised video person re-id methods on three benchmarks: PRID 2011, iLIDS-VID and MARS.
We consider the problem of person search in unconstrained scene images. Existing methods usually focus on improving the person detection accuracy to mitigate negative effects imposed by misalignment, mis-detections, and false alarms resulted from noisy people auto-detection. In contrast to previous studies, we show that sufficiently reliable person instance cropping is achievable by slightly improved state-of-the-art deep learning object detectors (e.g. Faster-RCNN), and the under-studied multi-scale matching problem in person search is a more severe barrier. In this work, we address this multi-scale person search challenge by proposing a Cross-Level Semantic Alignment (CLSA) deep learning approach capable of learning more discriminative identity feature representations in a unified end-to-end model. This is realised by exploiting the in-network feature pyramid structure of a deep neural network enhanced by a novel cross pyramid-level semantic alignment loss function. This favourably eliminates the need for constructing a computationally expensive image pyramid and a complex multi-branch network architecture. Extensive experiments show the modelling advantages and performance superiority of CLSA over the state-of-the-art person search and multi-scale matching methods on two large person search benchmarking datasets: CUHK-SYSU and PRW.
Existing person re-identification (re-id) methods assume the provision of accurately cropped person bounding boxes with minimum background noise, mostly by manually cropping. This is significantly breached in practice when person bounding boxes must be detected automatically given a very large number of images and/or videos processed. Compared to carefully cropped manually, auto-detected bounding boxes are far less accurate with random amount of background clutter which can degrade notably person re-id matching accuracy. In this work, we develop a joint learning deep model that optimises person re-id attention selection within any auto-detected person bounding boxes by reinforcement learning of background clutter minimisation subject to re-id label pairwise constraints. Specifically, we formulate a novel unified re-id architecture called Identity DiscriminativE Attention reinforcement Learning (IDEAL) to accurately select re-id attention in auto-detected bounding boxes for optimising re-id performance. Our model can improve re-id accuracy comparable to that from exhaustive human manual cropping of bounding boxes with additional advantages from identity discriminative attention selection that specially benefits re-id tasks beyond human knowledge. Extensive comparative evaluations demonstrate the re-id advantages of the proposed IDEAL model over a wide range of state-of-the-art re-id methods on two auto-detected re-id benchmarks CUHK03 and Market-1501.