There has been an increasing interest in learning dynamics simulators for model-based control. Compared with off-the-shelf physics engines, a learnable simulator can quickly adapt to unseen objects, scenes, and tasks. However, existing models like interaction networks only work for fully observable systems; they also only consider pairwise interactions within a single time step, both restricting their use in practical systems. We introduce Propagation Networks (PropNet), a differentiable, learnable dynamics model that handles partially observable scenarios and enables instantaneous propagation of signals beyond pairwise interactions. With these innovations, our propagation networks not only outperform current learnable physics engines in forward simulation, but also achieves superior performance on various control tasks. Compared with existing deep reinforcement learning algorithms, model-based control with propagation networks is more accurate, efficient, and generalizable to novel, partially observable scenes and tasks.
We introduce a saliency-based distortion layer for convolutional neural networks that helps to improve the spatial sampling of input data for a given task. Our differentiable layer can be added as a preprocessing block to existing task networks and trained altogether in an end-to-end fashion. The effect of the layer is to efficiently estimate how to sample from the original data in order to boost task performance. For example, for an image classification task in which the original data might range in size up to several megapixels, but where the desired input images to the task network are much smaller, our layer learns how best to sample from the underlying high resolution data in a manner which preserves task-relevant information better than uniform downsampling. This has the effect of creating distorted, caricature-like intermediate images, in which idiosyncratic elements of the image that improve task performance are zoomed and exaggerated. Unlike alternative approaches such as spatial transformer networks, our proposed layer is inspired by image saliency, computed efficiently from uniformly downsampled data, and degrades gracefully to a uniform sampling strategy under uncertainty. We apply our layer to improve existing networks for the tasks of human gaze estimation and fine-grained object classification. Code for our method is available in: http://github.com/recasens/Saliency-Sampler
In this work, we introduce pose interpreter networks for 6-DoF object pose estimation. In contrast to other CNN-based approaches to pose estimation that require expensively annotated object pose data, our pose interpreter network is trained entirely on synthetic pose data. We use object masks as an intermediate representation to bridge real and synthetic. We show that when combined with a segmentation model trained on RGB images, our synthetically trained pose interpreter network is able to generalize to real data. Our end-to-end system for object pose estimation runs in real-time (20 Hz) on live RGB data, without using depth information or ICP refinement.
Temporal relational reasoning, the ability to link meaningful transformations of objects or entities over time, is a fundamental property of intelligent species. In this paper, we introduce an effective and interpretable network module, the Temporal Relation Network (TRN), designed to learn and reason about temporal dependencies between video frames at multiple time scales. We evaluate TRN-equipped networks on activity recognition tasks using three recent video datasets - Something-Something, Jester, and Charades - which fundamentally depend on temporal relational reasoning. Our results demonstrate that the proposed TRN gives convolutional neural networks a remarkable capacity to discover temporal relations in videos. Through only sparsely sampled video frames, TRN-equipped networks can accurately predict human-object interactions in the Something-Something dataset and identify various human gestures on the Jester dataset with very competitive performance. TRN-equipped networks also outperform two-stream networks and 3D convolution networks in recognizing daily activities in the Charades dataset. Further analyses show that the models learn intuitive and interpretable visual common sense knowledge in videos.
The success of recent deep convolutional neural networks (CNNs) depends on learning hidden representations that can summarize the important factors of variation behind the data. However, CNNs often criticized as being black boxes that lack interpretability, since they have millions of unexplained model parameters. In this work, we describe Network Dissection, a method that interprets networks by providing labels for the units of their deep visual representations. The proposed method quantifies the interpretability of CNN representations by evaluating the alignment between individual hidden units and a set of visual semantic concepts. By identifying the best alignments, units are given human interpretable labels across a range of objects, parts, scenes, textures, materials, and colors. The method reveals that deep representations are more transparent and interpretable than expected: we find that representations are significantly more interpretable than they would be under a random equivalently powerful basis. We apply the method to interpret and compare the latent representations of various network architectures trained to solve different supervised and self-supervised training tasks. We then examine factors affecting the network interpretability such as the number of the training iterations, regularizations, different initializations, and the network depth and width. Finally we show that the interpreted units can be used to provide explicit explanations of a prediction given by a CNN for an image. Our results highlight that interpretability is an important property of deep neural networks that provides new insights into their hierarchical structure.
In this paper, we are interested in modeling complex activities that occur in a typical household. We propose to use programs, i.e., sequences of atomic actions and interactions, as a high level representation of complex tasks. Programs are interesting because they provide a non-ambiguous representation of a task, and allow agents to execute them. However, nowadays, there is no database providing this type of information. Towards this goal, we first crowd-source programs for a variety of activities that happen in people's homes, via a game-like interface used for teaching kids how to code. Using the collected dataset, we show how we can learn to extract programs directly from natural language descriptions or from videos. We then implement the most common atomic (inter)actions in the Unity3D game engine, and use our programs to "drive" an artificial agent to execute tasks in a simulated household environment. Our VirtualHome simulator allows us to create a large activity video dataset with rich ground-truth, enabling training and testing of video understanding models. We further showcase examples of our agent performing tasks in our VirtualHome based on language descriptions.
We address the problem of affordance reasoning in diverse scenes that appear in the real world. Affordances relate the agent's actions to their effects when taken on the surrounding objects. In our work, we take the egocentric view of the scene, and aim to reason about action-object affordances that respect both the physical world as well as the social norms imposed by the society. We also aim to teach artificial agents why some actions should not be taken in certain situations, and what would likely happen if these actions would be taken. We collect a new dataset that builds upon ADE20k, referred to as ADE-Affordance, which contains annotations enabling such rich visual reasoning. We propose a model that exploits Graph Neural Networks to propagate contextual information from the scene in order to perform detailed affordance reasoning about each object. Our model is showcased through various ablation studies, pointing to successes and challenges in this complex task.
We revisit the importance of the individual units in Convolutional Neural Networks (CNNs) for visual recognition. By conducting unit ablation experiments on CNNs trained on large scale image datasets, we demonstrate that, though ablating any individual unit does not hurt overall classification accuracy, it does lead to significant damage on the accuracy of specific classes. This result shows that an individual unit is specialized to encode information relevant to a subset of classes. We compute the correlation between the accuracy drop under unit ablation and various attributes of an individual unit such as class selectivity and weight L1 norm. We confirm that unit attributes such as class selectivity are a poor predictor for impact on overall accuracy as found previously in recent work \cite{morcos2018importance}. However, our results show that class selectivity along with other attributes are good predictors of the importance of one unit to individual classes. We evaluate the impact of random rotation, batch normalization, and dropout to the importance of units to specific classes. Our results show that units with high selectivity play an important role in network classification power at the individual class level. Understanding and interpreting the behavior of these units is necessary and meaningful.
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.