In this paper, we explore neural network models that learn to associate segments of spoken audio captions with the semantically relevant portions of natural images that they refer to. We demonstrate that these audio-visual associative localizations emerge from network-internal representations learned as a by-product of training to perform an image-audio retrieval task. Our models operate directly on the image pixels and speech waveform, and do not rely on any conventional supervision in the form of labels, segmentations, or alignments between the modalities during training. We perform analysis using the Places 205 and ADE20k datasets demonstrating that our models implicitly learn semantically-coupled object and word detectors.
Understanding 3D object structure from a single image is an important but challenging task in computer vision, mostly due to the lack of 3D object annotations to real images. Previous research tackled this problem by either searching for a 3D shape that best explains 2D annotations, or training purely on synthetic data with ground truth 3D information. In this work, we propose 3D INterpreter Networks (3D-INN), an end-to-end trainable framework that sequentially estimates 2D keypoint heatmaps and 3D object skeletons and poses. Our system learns from both 2D-annotated real images and synthetic 3D data. This is made possible mainly by two technical innovations. First, heatmaps of 2D keypoints serve as an intermediate representation to connect real and synthetic data. 3D-INN is trained on real images to estimate 2D keypoint heatmaps from an input image; it then predicts 3D object structure from heatmaps using knowledge learned from synthetic 3D shapes. By doing so, 3D-INN benefits from the variation and abundance of synthetic 3D objects, without suffering from the domain difference between real and synthesized images, often due to imperfect rendering. Second, we propose a Projection Layer, mapping estimated 3D structure back to 2D. During training, it ensures 3D-INN to predict 3D structure whose projection is consistent with the 2D annotations to real images. Experiments show that the proposed system performs well on both 2D keypoint estimation and 3D structure recovery. We also demonstrate that the recovered 3D information has wide vision applications, such as image retrieval.
This paper describes a procedure for the creation of large-scale video datasets for action classification and localization from unconstrained, realistic web data. The scalability of the proposed procedure is demonstrated by building a novel video benchmark, named SLAC (Sparsely Labeled ACtions), consisting of over 520K untrimmed videos and 1.75M clip annotations spanning 200 action categories. Using our proposed framework, annotating a clip takes merely 8.8 seconds on average. This represents a saving in labeling time of over 95% compared to the traditional procedure of manual trimming and localization of actions. Our approach dramatically reduces the amount of human labeling by automatically identifying hard clips, i.e., clips that contain coherent actions but lead to prediction disagreement between action classifiers. A human annotator can disambiguate whether such a clip truly contains the hypothesized action in a handful of seconds, thus generating labels for highly informative samples at little cost. We show that our large-scale dataset can be used to effectively pre-train action recognition models, significantly improving final metrics on smaller-scale benchmarks after fine-tuning. On Kinetics, UCF-101 and HMDB-51, models pre-trained on SLAC outperform baselines trained from scratch, by 2.0%, 20.1% and 35.4% in top-1 accuracy, respectively when RGB input is used. Furthermore, we introduce a simple procedure that leverages the sparse labels in SLAC to pre-train action localization models. On THUMOS14 and ActivityNet-v1.3, our localization model improves the mAP of baseline model by 8.6% and 2.5%, respectively.
The sound of crashing waves, the roar of fast-moving cars -- sound conveys important information about the objects in our surroundings. In this work, we show that ambient sounds can be used as a supervisory signal for learning visual models. To demonstrate this, we train a convolutional neural network to predict a statistical summary of the sound associated with a video frame. We show that, through this process, the network learns a representation that conveys information about objects and scenes. We evaluate this representation on several recognition tasks, finding that its performance is comparable to that of other state-of-the-art unsupervised learning methods. Finally, we show through visualizations that the network learns units that are selective to objects that are often associated with characteristic sounds. This paper extends an earlier conference paper, Owens et al. 2016, with additional experiments and discussion.
Recent robotic manipulation competitions have highlighted that sophisticated robots still struggle to achieve fast and reliable perception of task-relevant objects in complex, realistic scenarios. To improve these systems' perceptive speed and robustness, we present SegICP, a novel integrated solution to object recognition and pose estimation. SegICP couples convolutional neural networks and multi-hypothesis point cloud registration to achieve both robust pixel-wise semantic segmentation as well as accurate and real-time 6-DOF pose estimation for relevant objects. Our architecture achieves 1cm position error and <5^\circ$ angle error in real time without an initial seed. We evaluate and benchmark SegICP against an annotated dataset generated by motion capture.
We capitalize on large amounts of readily-available, synchronous data to learn a deep discriminative representations shared across three major natural modalities: vision, sound and language. By leveraging over a year of sound from video and millions of sentences paired with images, we jointly train a deep convolutional network for aligned representation learning. Our experiments suggest that this representation is useful for several tasks, such as cross-modal retrieval or transferring classifiers between modalities. Moreover, although our network is only trained with image+text and image+sound pairs, it can transfer between text and sound as well, a transfer the network never observed during training. Visualizations of our representation reveal many hidden units which automatically emerge to detect concepts, independent of the modality.
We propose a general framework called Network Dissection for quantifying the interpretability of latent representations of CNNs by evaluating the alignment between individual hidden units and a set of semantic concepts. Given any CNN model, the proposed method draws on a broad data set of visual concepts to score the semantics of hidden units at each intermediate convolutional layer. The units with semantics are given labels across a range of objects, parts, scenes, textures, materials, and colors. We use the proposed method to test the hypothesis that interpretability of units is equivalent to random linear combinations of units, then we apply our method to compare the latent representations of various networks when trained to solve different supervised and self-supervised training tasks. We further analyze the effect of training iterations, compare networks trained with different initializations, examine the impact of network depth and width, and measure the effect of dropout and batch normalization on the interpretability of deep visual representations. We demonstrate that the proposed method can shed light on characteristics of CNN models and training methods that go beyond measurements of their discriminative power.
How best to evaluate a saliency model's ability to predict where humans look in images is an open research question. The choice of evaluation metric depends on how saliency is defined and how the ground truth is represented. Metrics differ in how they rank saliency models, and this results from how false positives and false negatives are treated, whether viewing biases are accounted for, whether spatial deviations are factored in, and how the saliency maps are pre-processed. In this paper, we provide an analysis of 8 different evaluation metrics and their properties. With the help of systematic experiments and visualizations of metric computations, we add interpretability to saliency scores and more transparency to the evaluation of saliency models. Building off the differences in metric properties and behaviors, we make recommendations for metric selections under specific assumptions and for specific applications.
Recognizing arbitrary objects in the wild has been a challenging problem due to the limitations of existing classification models and datasets. In this paper, we propose a new task that aims at parsing scenes with a large and open vocabulary, and several evaluation metrics are explored for this problem. Our proposed approach to this problem is a joint image pixel and word concept embeddings framework, where word concepts are connected by semantic relations. We validate the open vocabulary prediction ability of our framework on ADE20K dataset which covers a wide variety of scenes and objects. We further explore the trained joint embedding space to show its interpretability.
Recognizing when people have false beliefs is crucial for understanding their actions. We introduce the novel problem of identifying when people in abstract scenes have incorrect beliefs. We present a dataset of scenes, each visually depicting an 8-frame story in which a character has a mistaken belief. We then create a representation of characters' beliefs for two tasks in human action understanding: predicting who is mistaken, and when they are mistaken. Experiments suggest that our method for identifying mistaken characters performs better on these tasks than simple baselines. Diagnostics on our model suggest it learns important cues for recognizing mistaken beliefs, such as gaze. We believe models of people's beliefs will have many applications in action understanding, robotics, and healthcare.