ICCV 2017 Papers

We define the object detection from imagery problem as estimating a very large but extremely sparse bounding box dependent probability distribution. Subsequently we identify a sparse distribution estimation scheme, Directed Sparse Sampling, and employ it in a single end-to-end CNN based detection model. This methodology extends and formalizes previous state-of-the-art detection models with an additional emphasis on high evaluation rates and reduced manual engineering. We introduce two novelties, a corner based region-of-interest estimator and a deconvolution based CNN model. The resulting model is scene adaptive, does not require manually defined reference bounding boxes and produces highly competitive results on MSCOCO, Pascal VOC 2007 and Pascal VOC 2012 with real-time evaluation rates. Further analysis suggests our model performs particularly well when finegrained object localization is desirable. We argue that this advantage stems from the significantly larger set of available regions-of-interest relative to other methods. Source-code is available from: https://github.com/lachlants/denet

We present Deeply Supervised Object Detector (DSOD), a framework that can learn object detectors from scratch. State-of-the-art object objectors rely heavily on the off-the-shelf networks pre-trained on large-scale classification datasets like ImageNet, which incurs learning bias due to the difference on both the loss functions and the category distributions between classification and detection tasks. Model fine-tuning for the detection task could alleviate this bias to some extent but not fundamentally. Besides, transferring pre-trained models from classification to detection between discrepant domains is even more difficult (e.g. RGB to depth images). A better solution to tackle these two critical problems is to train object detectors from scratch, which motivates our proposed DSOD. Previous efforts in this direction mostly failed due to much more complicated loss functions and limited training data in object detection. In DSOD, we contribute a set of design principles for training object detectors from scratch. One of the key findings is that deep supervision, enabled by dense layer-wise connections, plays a critical role in learning a good detector. Combining with several other principles, we develop DSOD following the single-shot detection (SSD) framework. Experiments on PASCAL VOC 2007, 2012 and MS COCO datasets demonstrate that DSOD can achieve better results than the state-of-the-art solutions with much more compact models. For instance, DSOD outperforms SSD on all three benchmarks with real-time detection speed, while requires only 1/2 parameters to SSD and 1/10 parameters to Faster RCNN.

Despite a rapid rise in the quality of built-in smartphone cameras, their physical limitations - small sensor size, compact lenses and the lack of specific hardware, - impede them to achieve the quality results of DSLR cameras. In this work we present an end-to-end deep learning approach that bridges this gap by translating ordinary photos into DSLR-quality images. We propose learning the translation function using a residual convolutional neural network that improves both color rendition and image sharpness. Since the standard mean squared loss is not well suited for measuring perceptual image quality, we introduce a composite perceptual error function that combines content, color and texture losses. The first two losses are defined analytically, while the texture loss is learned in an adversarial fashion. We also present DPED, a large-scale dataset that consists of real photos captured from three different phones and one high-end reflex camera. Our quantitative and qualitative assessments reveal that the enhanced image quality is comparable to that of DSLR-taken photos, while the methodology is generalized to any type of digital camera.

Techniques for dense semantic correspondence have provided limited ability to deal with the geometric variations that commonly exist between semantically similar images. While variations due to scale and rotation have been examined, there is a lack of practical solutions for more complex deformations such as affine transformations because of the tremendous size of the associated solution space. To address this problem, we present a discrete-continuous transformation matching (DCTM) framework where dense affine transformation fields are inferred through a discrete label optimization in which the labels are iteratively updated via continuous regularization. In this way, our approach draws solutions from the continuous space of affine transformations in a manner that can be computed efficiently through constant-time edge-aware filtering and a proposed affine-varying CNN-based descriptor. Experimental results show that this model outperforms the state-of-the-art methods for dense semantic correspondence on various benchmarks.

A major impediment in rapidly deploying object detection models for instance detection is the lack of large annotated datasets. For example, finding a large labeled dataset containing instances in a particular kitchen is unlikely. Each new environment with new instances requires expensive data collection and annotation. In this paper, we propose a simple approach to generate large annotated instance datasets with minimal effort. Our key insight is that ensuring only patch-level realism provides enough training signal for current object detector models. We automatically `cut' object instances and `paste' them on random backgrounds. A naive way to do this results in pixel artifacts which result in poor performance for trained models. We show how to make detectors ignore these artifacts during training and generate data that gives competitive performance on real data. Our method outperforms existing synthesis approaches and when combined with real images improves relative performance by more than 21% on benchmark datasets. In a cross-domain setting, our synthetic data combined with just 10% real data outperforms models trained on all real data.

Dropout is a very effective way of regularizing neural networks. Stochastically dropping out units with a certain probability discourages over-specific co-adaptations of feature detectors, preventing overfitting and improving network generalization. Besides, Dropout can be interpreted as an approximate model aggregation technique, where an exponential number of smaller networks are averaged in order to get a more powerful ensemble. In this paper, we show that using a fixed dropout probability during training is a suboptimal choice. We thus propose a time scheduling for the probability of retaining neurons in the network. This induces an adaptive regularization scheme that smoothly increases the difficulty of the optimization problem. This idea of starting easy and adaptively increasing the difficulty of the learning problem has its roots in curriculum learning and allows one to train better models. Indeed, we prove that our optimization strategy implements a very general curriculum scheme, by gradually adding noise to both the input and intermediate feature representations within the network architecture. Experiments on seven image classification datasets and different network architectures show that our method, named Curriculum Dropout, frequently yields to better generalization and, at worst, performs just as well as the standard Dropout method.

During the last half decade, convolutional neural networks (CNNs) have triumphed over semantic segmentation, which is a core task of various emerging industrial applications such as autonomous driving and medical imaging. However, to train CNNs requires a huge amount of data, which is difficult to collect and laborious to annotate. Recent advances in computer graphics make it possible to train CNN models on photo-realistic synthetic data with computer-generated annotations. Despite this, the domain mismatch between the real images and the synthetic data significantly decreases the models' performance. Hence we propose a curriculum-style learning approach to minimize the domain gap in semantic segmentation. The curriculum domain adaptation solves easy tasks first in order to infer some necessary properties about the target domain; in particular, the first task is to learn global label distributions over images and local distributions over landmark superpixels. These are easy to estimate because images of urban traffic scenes have strong idiosyncrasies (e.g., the size and spatial relations of buildings, streets, cars, etc.). We then train the segmentation network in such a way that the network predictions in the target domain follow those inferred properties. In experiments, our method significantly outperforms the baselines as well as the only known existing approach to the same problem.

While metric learning is important for Person re-identification (RE-ID), a significant problem in visual surveillance for cross-view pedestrian matching, existing metric models for RE-ID are mostly based on supervised learning that requires quantities of labeled samples in all pairs of camera views for training. However, this limits their scalabilities to realistic applications, in which a large amount of data over multiple disjoint camera views is available but not labelled. To overcome the problem, we propose an unsupervised asymmetric metric learning model for unsupervised RE-ID. Our model aims to learn an asymmetric metric, i.e., specific projection for each view, effectively based on clustering on cross-view person images. Our model finds a shared space where view-specific bias is alleviated and thus better matching performance can be achieved. Extensive experiments have been conducted on a baseline and five large-scale RE-ID datasets to demonstrate the effectiveness of the proposed model. Through the comparison, we show that our unsupervised asymmetric metric model works much more suitable for unsupervised RE-ID as compared to classical unsupervised metric learning models. We also compare existing unsupervised RE-ID methods, and our model outperforms them with notable margins, and especially we report the performance on large-scale unlabelled RE-ID dataset, which is unfortunately less concerned in literatures.

In this paper, we propose a cross-modal deep variational hashing (CMDVH) method to learn compact binary codes for cross-modality multimedia retrieval. Unlike most existing cross-modal hashing methods which learn a single pair of projections to map each example into a binary vector, we design a deep fusion neural network to learn non-linear transformations from image-text input pairs, such that a unified binary code is achieved in a discrete and discriminative manner using a classification-based hinge-loss criterion. We then design modality-specific neural networks in a probabilistic manner such that we model a latent variable to be close as possible from the inferred binary codes, at the same time approximated by a posterior distribution regularized by a known prior, which is suitable for out-of-sample extension. Experimental results on three benchmark datasets show the efficacy of the proposed approach.

The region-based Convolutional Neural Network (CNN) detectors such as Faster R-CNN or R-FCN have already shown promising results for object detection by combining the region proposal subnetwork and the classification subnetwork together. Although R-FCN has achieved higher detection speed while keeping the detection performance, the global structure information is ignored by the position-sensitive score maps. To fully explore the local and global properties, in this paper, we propose a novel fully convolutional network, named as CoupleNet, to couple the global structure with local parts for object detection. Specifically, the object proposals obtained by the Region Proposal Network (RPN) are fed into the the coupling module which consists of two branches. One branch adopts the position-sensitive RoI (PSRoI) pooling to capture the local part information of the object, while the other employs the RoI pooling to encode the global and context information. Next, we design different coupling strategies and normalization ways to make full use of the complementary advantages between the global and local branches. Extensive experiments demonstrate the effectiveness of our approach. We achieve state-of-the-art results on all three challenging datasets, i.e. a mAP of 82.7% on VOC07, 80.4% on VOC12, and 34.4% on COCO.Codes will be made publicly available.

We propose a method for geometric calibration of multi-focus plenoptic cameras using raw images. Multi-focus plenoptic cameras feature several types of micro-lenses spatially aligned in front of the camera sensor to generate micro-images at different magnifications. This multi-lens arrangement provides computational-photography benefits but complicates calibration. Our methodology achieves the detection of the type of micro-lenses, the retrieval of their spatial arrangement, and the estimation of intrinsic and extrinsic camera parameters therefore fully characterising this specialised camera class. Motivated from classic pinhole camera calibration, the presented algorithm operates based on a checker-board's corners, retrieved by a custom micro-image corner detector. This approach enables the introduction of a re-projection error that is used in a minimisation framework. Our algorithm compares favourably to the state-of-the-art, as demonstrated by controlled and free-hand experiments, making it a first step towards accurate 3D reconstruction and Structure-from-Motion.

Very large-scale Deep Neural Networks (DNNs) have achieved remarkable successes in a large variety of computer vision tasks. However, the high computation intensity of DNNs makes it challenging to deploy these models on resource-limited systems. Some studies used low-rank approaches that approximate the filters by low-rank basis to accelerate the testing. Those works directly decomposed the pre-trained DNNs by Low-Rank Approximations (LRA). How to train DNNs toward lower-rank space for more efficient DNNs, however, remains as an open area. To solve the issue, in this work, we propose Force Regularization, which uses attractive forces to enforce filters so as to coordinate more weight information into lower-rank space. We mathematically and empirically verify that after applying our technique, standard LRA methods can reconstruct filters using much lower basis and thus result in faster DNNs. The effectiveness of our approach is comprehensively evaluated in ResNets, AlexNet, and GoogLeNet. In AlexNet, for example, Force Regularization gains 2x speedup on modern GPU without accuracy loss and 4.05x speedup on CPU by paying small accuracy degradation. Moreover, Force Regularization better initializes the low-rank DNNs such that the fine-tuning can converge faster toward higher accuracy. The obtained lower-rank DNNs can be further sparsified, proving that Force Regularization can be integrated with state-of-the-art sparsity-based acceleration methods.

Convolutional Sparse Coding (CSC) is an increasingly popular model in the signal and image processing communities, tackling some of the limitations of traditional patch-based sparse representations. Although several works have addressed the dictionary learning problem under this model, these relied on an ADMM formulation in the Fourier domain, losing the sense of locality and the relation to the traditional patch-based sparse pursuit. A recent work suggested a novel theoretical analysis of this global model, providing guarantees that rely on a localized sparsity measure. Herein, we extend this local-global relation by showing how one can efficiently solve the convolutional sparse pursuit problem and train the filters involved, while operating locally on image patches. Our approach provides an intuitive algorithm that can leverage standard techniques from the sparse representations field. The proposed method is fast to train, simple to implement, and flexible enough that it can be easily deployed in a variety of applications. We demonstrate the proposed training scheme for image inpainting and image separation, while achieving state-of-the-art results.

One popular approach for blind deconvolution is to formulate a maximum a posteriori (MAP) problem with sparsity priors on the gradients of the latent image, and then alternatingly estimate the blur kernel and the latent image. While several successful MAP based methods have been proposed, there has been much controversy and confusion about their convergence, because sparsity priors have been shown to prefer blurry images to sharp natural images. In this paper, we revisit this problem and provide an analysis on the convergence of MAP based approaches. We first introduce a slight modification to a conventional joint energy function for blind deconvolution. The reformulated energy function yields the same alternating estimation process, but more clearly reveals how blind deconvolution works. We then show the energy function can actually favor the right solution instead of the no-blur solution under certain conditions, which explains the success of previous MAP based approaches. The reformulated energy function and our conditions for the convergence also provide a way to compare the qualities of different blur kernels, and we demonstrate its applicability to automatic blur kernel size selection, blur kernel estimation using light streaks, and defocus estimation.

Convolutional sparse coding (CSC) plays an essential role in many computer vision applications ranging from image compression to deep learning. In this work, we spot the light on a new application where CSC can effectively serve, namely line drawing analysis. The process of drawing a line drawing can be approximated as the sparse spatial localization of a number of typical basic strokes, which in turn can be cast as a non-standard CSC model that considers the line drawing formation process from parametric curves. These curves are learned to optimize the fit between the model and a specific set of line drawings. Parametric representation of sketches is vital in enabling automatic sketch analysis, synthesis and manipulation. A couple of sketch manipulation examples are demonstrated in this work. Consequently, our novel method is expected to provide a reliable and automatic method for parametric sketch description. Through experiments, we empirically validate the convergence of our method to a feasible solution.

Convolutional sparse coding (CSC) is a promising direction for unsupervised learning in computer vision. In contrast to recent supervised methods, CSC allows for convolutional image representations to be learned that are equally useful for high-level vision tasks and low-level image reconstruction and can be applied to a wide range of tasks without problem-specific retraining. Due to their extreme memory requirements, however, existing CSC solvers have so far been limited to low-dimensional problems and datasets using a handful of low-resolution example images at a time. In this paper, we propose a new approach to solving CSC as a consensus optimization problem, which lifts these limitations. By learning CSC features from large-scale image datasets for the first time, we achieve significant quality improvements in a number of imaging tasks. Moreover, the proposed method enables new applications in high-dimensional feature learning that has been intractable using existing CSC methods. This is demonstrated for a variety of reconstruction problems across diverse problem domains, including 3D multispectral demosaicing and 4D light field view synthesis.

Most of current visual search systems focus on image-to-image (point-to-point) search such as image and object retrieval. Nevertheless, fast image-to-video (point-to-set) search is much less exploited. This paper tackles object instance search in videos, where efficient point-to-set matching is essential. Through jointly optimizing vector quantization and hashing, we propose compressive quantization method to compress M object proposals extracted from each video into only k binary codes, where k<< M. Then the similarity between the query object and the whole video can be determined by the Hamming distance between the query's binary code and the video's best-matched binary code. Our compressive quantization not only enables fast search but also significantly reduces the memory cost of storing the video features. Despite the high compression ratio, our proposed compressive quantization still can effectively retrieve small objects in large video datasets. Systematic experiments on three benchmark datasets verify the effectiveness and efficiency of our compressive quantization.

Regression based methods are not performing as well as detection based methods for human pose estimation. A central problem is that the structural information in the pose is not well exploited in the previous regression methods. In this work, we propose a structure-aware regression approach. It adopts a reparameterized pose representation using bones instead of joints. It exploits the joint connection structure to define a compositional loss function that encodes the long range interactions in the pose. It is simple, effective, and general for both 2D and 3D pose estimation in a unified setting. Comprehensive evaluation validates the effectiveness of our approach. It significantly advances the state-of-the-art on Human3.6M and is competitive with state-of-the-art results on MPII.

Depth from focus is a highly accessible method to estimate the 3D structure of everyday scenes. Today's DSLR and mobile cameras facilitate the easy capture of multiple focused images of a scene. Focus measures (FMs) that estimate the amount of focus at each pixel form the basis of depth-from-focus methods. Several FMs have been proposed in the past and new ones will emerge in the future, each with their own strengths. We estimate a weighted combination of standard FMs that outperforms others on a wide range of scene types. The resulting composite focus measure consists of FMs that are in consensus with one another but not in chorus. Our two-stage pipeline first estimates fine depth at each pixel using the composite focus measure. A cost-volume propagation step then assigns depths from confident pixels to others. We can generate high quality depth maps using just the top five FMs from our composite focus measure. This is a positive step towards depth estimation of everyday scenes with no special equipment.