CVPR 2016 Papers

In this paper, we introduce a novel framework for WEakly supervised Learning of Deep cOnvolutional neural Networks (WELDON). Our method is dedicated to automatically selecting relevant image regions from weak annotations, e.g. global image labels, and encompasses the following contributions. Firstly, WELDON leverages recent improvements on the Multiple Instance Learning paradigm, i.e. negative evidence scoring and top instance selection. Secondly, the deep CNN is trained to optimize Average Precision, and fine-tuned on the target dataset with efficient computations due to convolutional feature sharing. A thorough experimental validation shows that WELDON outperforms state-of-the-art results on six different datasets.

3D shape models are becoming widely available and easier to capture, making available 3D information crucial for progress in object classification. Current state-of-the-art methods rely on CNNs to address this problem. Recently, we witness two types of CNNs being developed: CNNs based upon volumetric representations versus CNNs based upon multi-view representations. Empirical results from these two types of CNNs exhibit a large gap, indicating that existing volumetric CNN architectures and approaches are unable to fully exploit the power of 3D representations. In this paper, we aim to improve both volumetric CNNs and multi-view CNNs according to extensive analysis of existing approaches. To this end, we introduce two distinct network architectures of volumetric CNNs. In addition, we examine multi-view CNNs, where we introduce multi-resolution filtering in 3D. Overall, we are able to outperform current state-of-the-art methods for both volumetric CNNs and multi-view CNNs. We provide extensive experiments designed to evaluate underlying design choices, thus providing a better understanding of the space of methods available for object classification on 3D data.

In this paper, we propose a new framework for 3D tracking by detection based on fully volumetric representations. On one hand, 3D tracking by detection has shown robust use in the context of interaction (Kinect) and surface tracking. On the other hand, volumetric representations have recently been proven efficient both for building 3D features and for addressing the 3D tracking problem. We leverage these benefits by unifying both families of approaches into a single, fully volumetric tracking-by-detection framework. We use a centroidal Voronoi tessellation (CVT) representation to compactly tessellate shapes with optimal discretization, construct a feature space, and perform the tracking according to the correspondences provided by trained random forests. Our results show improved tracking and training computational efficiency and improved memory performance. This in turn enables the use of larger training databases than state of the art approaches, which we leverage by proposing a cross-tracking subject training scheme to benefit from all subject sequences for all tracking situations, thus yielding better detection and less overfitting.

Objects make distinctive sounds when they are hit or scratched. These sounds reveal aspects of an object's material properties, as well as the actions that produced them. In this paper, we propose the task of predicting what sound an object makes when struck as a way of studying physical interactions within a visual scene. We present an algorithm that synthesizes sound from silent videos of people hitting and scratching objects with a drumstick. This algorithm uses a recurrent neural network to predict sound features from videos and then produces a waveform from these features with an example-based synthesis procedure. We show that the sounds predicted by our model are realistic enough to fool participants in a "real or fake" psychophysical experiment, and that they convey significant information about material properties and physical interactions.

A number of recent approaches have used deep convolutional neural networks (CNNs) to build texture representations. Nevertheless, it is still unclear how these mod- els represent texture and invariances to categorical variations. This work conducts a systematic evaluation of recent CNN-based texture descriptors for recognition and attempts to understand the nature of invariances captured by these representations. First we show that the recently proposed bilinear CNN model [25] is an excellent generalpurpose texture descriptor and compares favorably to other CNN-based descriptors on various texture and scene recognition benchmarks. The model is translationally invariant and obtains better accuracy on the ImageNet dataset without requiring spatial jittering of data compared to corresponding models trained with spatial jittering. Based on recent work [13, 28] we propose a technique to visualize pre-images, providing a means for understanding categorical properties that are captured by these representations. Finally, we show preliminary results on how a unified parametric model of texture analysis and synthesis can be used for attribute-based image manipulation, e.g. to make an image more swirly, honeycombed, or knitted. The source code and additional visualizations are available at http://vis-www.cs.umass.edu/texture.

We have seen great progress in basic perceptual tasks such as object recognition and detection. However, AI models still fail to match humans in high-level vision tasks due to the lack of capacities for deeper reasoning. Recently the new task of visual question answering (QA) has been proposed to evaluate a model's capacity for deep image understanding. Previous works have established a loose, global association between QA sentences and images. However, many questions and answers, in practice, relate to local regions in the images. We establish a semantic link between textual descriptions and image regions by object-level grounding. It enables a new type of QA with visual answers, in addition to textual answers used in previous work. We study the visual QA tasks in a grounded setting with a large collection of 7W multiple-choice QA pairs. Furthermore, we evaluate human performance and several baseline models on the QA tasks. Finally, we propose a novel LSTM model with spatial attention to tackle the 7W QA tasks.

We propose a model to learn visually grounded word embeddings (vis-w2v) to capture visual notions of semantic relatedness. While word embeddings trained using text have been extremely successful, they cannot uncover notions of semantic relatedness implicit in our visual world. For instance, although "eats" and "stares at" seem unrelated in text, they share semantics visually. When people are eating something, they also tend to stare at the food. Grounding diverse relations like "eats" and "stares at" into vision remains challenging, despite recent progress in vision. We note that the visual grounding of words depends on semantics, and not the literal pixels. We thus use abstract scenes created from clipart to provide the visual grounding. We find that the embeddings we learn capture fine-grained, visually grounded notions of semantic relatedness. We show improvements over text-only word embeddings (word2vec) on three tasks: common-sense assertion classification, visual paraphrasing and text-based image retrieval. Our code and datasets are available online.

In this paper, we present a novel attention-modulated visual tracking algorithm that decomposes an object into multiple cognitive units, and trains multiple elementary trackers in order to modulate the distribution of attention according to various feature and kernel types. In the integration stage it recombines the units to memorize and recognize the target object effectively. With respect to the elementary trackers, we present a novel attentional feature-based correlation filter (AtCF) that focuses on distinctive attentional features. The effectiveness of the proposed algorithm is validated through experimental comparison with state-of-the-art methods on widely-used tracking benchmark datasets.

This paper proposes a novel path prediction algorithm for progressing one step further than the existing works focusing on single target path prediction. In this paper, we consider moving dynamics of co-occurring objects for path prediction in a scene that includes crowded moving objects. To solve this problem, we first suggest a two-layered probabilistic model to find major movement patterns and their co-occurrence tendency. By utilizing the unsupervised learning results from the model, we present an algorithm to find the future location of any target object. Through extensive qualitative/quantitative experiments, we show that our algorithm can find a plausible future path in complex scenes with a large number of moving objects.

Modern computer vision algorithms typically require expensive data acquisition and accurate manual labeling. In this work, we instead leverage the recent progress in computer graphics to generate fully labeled, dynamic, and photo-realistic proxy virtual worlds. We propose an efficient real-to-virtual world cloning method, and validate our approach by building and publicly releasing a new video dataset, called Virtual KITTI, automatically labeled with accurate ground truth for object detection, tracking, scene and instance segmentation, depth, and optical flow. We provide quantitative experimental evidence suggesting that (i) modern deep learning algorithms pre-trained on real data behave similarly in real and virtual worlds, and (ii) pre-training on virtual data improves performance. As the gap between real and virtual worlds is small, virtual worlds enable measuring the impact of various weather and imaging conditions on recognition performance, all other things being equal. We show these factors may affect drastically otherwise high-performing deep models for tracking.

We introduce the novel problem of automatically generating animated GIFs from video. GIFs are short looping video with no sound, and a perfect combination between image and video that really capture our attention. GIFs tell a story, express emotion, turn events into humorous moments, and are the new wave of photojournalism. We pose the question: Can we automate the entirely manual and elaborate process of GIF creation by leveraging the plethora of user generated GIF content? We propose a Robust Deep RankNet that, given a video, generates a ranked list of its segments according to their suitability as GIF. We train our model to learn what visual content is often selected for GIFs by using over 100K user generated GIFs and their corresponding video sources. We effectively deal with the noisy web data by proposing a novel adaptive Huber loss in the ranking formulation. We show that our approach is robust to outliers and picks up several patterns that are frequently present in popular animated GIFs. On our new large-scale benchmark dataset, we show the advantage of our approach over several state-of-the-art methods.

We address the problem of composing a story out of multiple short video clips taken by a person during an activity or experience. Inspired by plot analysis of written stories, our method generates a sequence of video clips ordered in such a way that it reflects plot dynamics and content coherency. That is, given a set of multiple video clips, our method composes a video which we call a video-story. We define metrics on scene dynamics and coherency by dense optical flow features and a patch matching algorithm. Using these metrics, we define an objective function for the video-story. To efficiently search for the best video-story, we introduce a novel Branch-and-Bound algorithm which guarantees the global optimum. We collect the dataset consisting of 23 video sets from the web, resulting in a total of 236 individual video clips. With the acquired dataset, we perform extensive user studies involving 30 human subjects by which the effectiveness of our approach is quantitatively and qualitatively verified.

Video object segmentation is challenging due to fast moving objects, deforming shapes, and cluttered backgrounds. Optical flow can be used to propagate an object segmentation over time but, unfortunately, flow is often inaccurate, particularly around object boundaries. Such boundaries are precisely where we want our segmentation to be accurate. To obtain accurate segmentation across time, we propose an efficient algorithm that considers video segmentation and optical flow estimation simultaneously. For video segmentation, we formulate a principled, multi-scale, spatio-temporal objective function that uses optical flow to propagate information between frames. For optical flow estimation, particularly at object boundaries, we compute the flow independently in the segmented regions and recompose the results. We call the process object flow and demonstrate the effectiveness of jointly optimizing optical flow and video segmentation using an iterative scheme. Experiments on the SegTrack v2 and Youtube-Objects datasets show that the proposed algorithm performs favorably against the other state-of-the-art methods.

We present an approach that exploits hierarchical Recurrent Neural Networks (RNNs) to tackle the video captioning problem, i.e., generating one or multiple sentences to describe a realistic video. Our hierarchical framework contains a sentence generator and a paragraph generator. The sentence generator produces one simple short sentence that describes a specific short video interval. It exploits both temporal- and spatial-attention mechanisms to selectively focus on visual elements during generation. The paragraph generator captures the inter-sentence dependency by taking as input the sentential embedding produced by the sentence generator, combining it with the paragraph history, and outputting the new initial state for the sentence generator. We evaluate our approach on two large-scale benchmark datasets: YouTubeClips and TACoS-MultiLevel. The experiments demonstrate that our approach significantly outperforms the current state-of-the-art methods with BLEU@4 scores 0.499 and 0.305 respectively.

Light fields have many applications in machine vision, consumer photography, robotics, and microscopy. However, the prevalent resolution limits of existing light field imaging systems hinder widespread adoption. In this paper, we analyze fundamental resolution limits of light field cameras in the diffraction limit. We propose a sequential, coded-aperture-style acquisition scheme that optimizes the resolution of a light field reconstructed from multiple photographs captured from different perspectives and f-number settings. We also show that the proposed acquisition scheme facilitates high dynamic range light field imaging and demonstrate a proof-of-concept prototype system. With this work, we hope to advance our understanding of the resolution limits of light field photography and develop practical computational imaging systems to overcome them.

Previous approaches to action recognition with deep features tend to process video frames only within a small temporal region, and do not model long-range dynamic information explicitly. However, such information is important for the accurate recognition of actions, especially for the discrimination of complex activities that share sub-actions, and when dealing with untrimmed videos. Here, we propose a representation, VLAD for Deep Dynamics (VLAD^3), that accounts for different levels of video dynamics. It captures short-term dynamics with deep convolutional neural network features, relying on linear dynamic systems (LDS) to model medium-range dynamics. To account for long-range inhomogeneous dynamics, a VLAD descriptor is derived for the LDS and pooled over the whole video, to arrive at the final VLAD^3 representation. An extensive evaluation was performed on Olympic Sports, UCF101 and THUMOS15, where the use of the VLAD^3 representation leads to state-of- the-art results.

Correctly matching feature points in a pair of images is an important preprocessing step for many computer vision applications. In this paper we propose an efficient method for estimating the number of correct matches without explicitly computing them. In addition, our method estimates the region of overlap between the images. To this end, we propose to analyze the set of matches using the spatial order of the features, as projected to the x-axis of the image. The set of features in each image is thus represented by a sequence. This reduces the analysis of the matching problem to the analysis of the permutation between the sequences. Using the Kendall distance metric between permutations and natural assumptions on the distribution of the correct and incorrect matches, we show how to estimate the above-mentioned values. We demonstrate the usefulness of our method in two applications: (i) a new halting condition for RANSAC based epipolar geometry estimation methods that considerably reduce the running time, and (ii) discarding spatially unrelated image pairs in the Structure-from-Motion pipeline. Furthermore, our analysis allows to compute the probability that a given match is correct based on the estimated number of correct matches and the rank of the features within the sequences. Our experiments on a large number of synthetic and real data demonstrate the effectiveness of our method. For example, the running time of the image matching stage in the Structure-from-Motion pipeline may be reduced by about 99% while preserving about 80% of the correctly matched feature points.

Learned confidence measures gain increasing importance for outlier removal and quality improvement in stereo vision. However, acquiring the necessary training data is typically a tedious and time consuming task that involves manual interaction, active sensing devices and/or synthetic scenes. To overcome this problem, we propose a new, flexible, and scalable way for generating training data that only requires a set of stereo images as input. The key idea of our approach is to use different view points for reasoning about contradictions and consistencies between multiple depth maps generated with the same stereo algorithm. This enables us to generate a huge amount of training data in a fully automated manner. Among other experiments, we demonstrate the potential of our approach by boosting the performance of three learned confidence measures on the KITTI2012 dataset by simply training them on a vast amount of automatically generated training data rather than a limited amount of laser ground truth data.

Data-driven approaches for edge detection have proven effective and achieve top results on modern benchmarks. However, all current data-driven edge detectors require manual supervision for training in the form of hand-labeled region segments or object boundaries. Specifically, human annotators mark semantically meaningful edges which are subsequently used for training. Is this form of strong, high-level supervision actually necessary to learn to accurately detect edges? In this work we present a simple yet effective approach for training edge detectors without human supervision. To this end we utilize motion, and more specifically, the only input to our method is noisy semi-dense matches between frames. We begin with only a rudimentary knowledge of edges (in the form of image gradients), and alternate between improving motion estimation and edge detection in turn. Using a large corpus of video data, we show that edge detectors trained using our unsupervised scheme approach the performance of the same methods trained with full supervision (within 3-5%). Finally, we show that when using a deep network for the edge detector, our approach provides a novel pre-training scheme for object detection.