This work explores the feasibility of steering a drone with a (recurrent) neural network, based on input from a forward looking camera, in the context of a high-level navigation task. We set up a generic framework for training a network to perform navigation tasks based on imitation learning. It can be applied to both aerial and land vehicles. As a proof of concept we apply it to a UAV (Unmanned Aerial Vehicle) in a simulated environment, learning to cross a room containing a number of obstacles. So far only feedforward neural networks (FNNs) have been used to train UAV control. To cope with more complex tasks, we propose the use of recurrent neural networks (RNN) instead and successfully train an LSTM (Long-Short Term Memory) network for controlling UAVs. Vision based control is a sequential prediction problem, known for its highly correlated input data. The correlation makes training a network hard, especially an RNN. To overcome this issue, we investigate an alternative sampling method during training, namely window-wise truncated backpropagation through time (WW-TBPTT). Further, end-to-end training requires a lot of data which often is not available. Therefore, we compare the performance of retraining only the Fully Connected (FC) and LSTM control layers with networks which are trained end-to-end. Performing the relatively simple task of crossing a room already reveals important guidelines and good practices for training neural control networks. Different visualizations help to explain the behavior learned.
In this work, we aim at automatically labeling actors in a TV series. Rather than relying on transcripts and subtitles, as has been demonstrated in the past, we show how to achieve this goal starting from a set of example images of each of the main actors involved, collected from the Internet Movie Database (IMDB). The problem then becomes one of domain adaptation: actors' IMDB photos are typically taken at awards ceremonies and are quite different from their appearances in TV series. In each series as well, there is considerable change in actor appearance due to makeup, lighting, ageing, etc. To bridge this gap, we propose a graph-matching based self-labelling algorithm, which we coin HSL (Hungarian Self Labeling). Further, we propose a new edge cost to be used in this context, as well as an extension that is more robust to outliers, where prototypical faces for each of the actors are selected based on a hierarchical clustering procedure. We conduct experiments with 15 episodes from 3 different TV series and demonstrate automatic annotation with an accuracy of 90% and up.
In this paper we focus on improving object detection performance in terms of recall. We propose a post-detection stage during which we explore the image with the objective of recovering missed detections. This exploration is performed by sampling object proposals in the image. We analyze four different strategies to perform this sampling, giving special attention to strategies that exploit spatial relations between objects. In addition, we propose a novel method to discover higher-order relations between groups of objects. Experiments on the challenging KITTI dataset show that our proposed relations-based proposal generation strategies can help improving recall at the cost of a relatively low amount of object proposals.
In online action detection, the goal is to detect the start of an action in a video stream as soon as it happens. For instance, if a child is chasing a ball, an autonomous car should recognize what is going on and respond immediately. This is a very challenging problem for four reasons. First, only partial actions are observed. Second, there is a large variability in negative data. Third, the start of the action is unknown, so it is unclear over what time window the information should be integrated. Finally, in real world data, large within-class variability exists. This problem has been addressed before, but only to some extent. Our contributions to online action detection are threefold. First, we introduce a realistic dataset composed of 27 episodes from 6 popular TV series. The dataset spans over 16 hours of footage annotated with 30 action classes, totaling 6,231 action instances. Second, we analyze and compare various baseline methods, showing this is a challenging problem for which none of the methods provides a good solution. Third, we analyze the change in performance when there is a variation in viewpoint, occlusion, truncation, etc. We introduce an evaluation protocol for fair comparison. The dataset, the baselines and the models will all be made publicly available to encourage (much needed) further research on online action detection on realistic data.
Taking an image of an object is at its core a lossy process. The rich information about the three-dimensional structure of the world is flattened to an image plane and decisions such as viewpoint and camera parameters are final and not easily revertible. As a consequence, possibilities of changing viewpoint are limited. Given a single image depicting an object, novel-view synthesis is the task of generating new images that render the object from a different viewpoint than the one given. The main difficulty is to synthesize the parts that are disoccluded; disocclusion occurs when parts of an object are hidden by the object itself under a specific viewpoint. In this work, we show how to improve novel-view synthesis by making use of the correlations observed in 3D models and applying them to new image instances. We propose a technique to use the structural information extracted from a 3D model that matches the image object in terms of viewpoint and shape. For the latter part, we propose an efficient 2D-to-3D alignment method that associates precisely the image appearance with the 3D model geometry with minimal user interaction. Our technique is able to simulate plausible viewpoint changes for a variety of object classes within seconds. Additionally, we show that our synthesized images can be used as additional training data that improves the performance of standard object detectors.
Over the years, hand gesture recognition has been mostly addressed considering hand trajectories in isolation. However, in most sign languages, hand gestures are defined on a particular context (body region). We propose a pipeline to perform sign language recognition which models hand movements in the context of other parts of the body captured in the 3D space using the MS Kinect sensor. In addition, we perform sign recognition based on the different hand postures that occur during a sign. Our experiments show that considering different body parts brings improved performance when compared to other methods which only consider global hand trajectories. Finally, we demonstrate that the combination of hand postures features with hand gestures features helps to improve the prediction of a given sign.
In this paper, a new method for generating object and action proposals in images and videos is proposed. It builds on activations of different convolutional layers of a pretrained CNN, combining the localization accuracy of the early layers with the high informative-ness (and hence recall) of the later layers. To this end, we build an inverse cascade that, going backward from the later to the earlier convolutional layers of the CNN, selects the most promising locations and refines them in a coarse-to-fine manner. The method is efficient, because i) it re-uses the same features extracted for detection, ii) it aggregates features using integral images, and iii) it avoids a dense evaluation of the proposals thanks to the use of the inverse coarse-to-fine cascade. The method is also accurate. We show that our DeepProposals outperform most of the previously proposed object proposal and action proposal approaches and, when plugged into a CNN-based object detector, produce state-of-the-art detection performance.
In a traditional convolutional layer, the learned filters stay fixed after training. In contrast, we introduce a new framework, the Dynamic Filter Network, where filters are generated dynamically conditioned on an input. We show that this architecture is a powerful one, with increased flexibility thanks to its adaptive nature, yet without an excessive increase in the number of model parameters. A wide variety of filtering operations can be learned this way, including local spatial transformations, but also others like selective (de)blurring or adaptive feature extraction. Moreover, multiple such layers can be combined, e.g. in a recurrent architecture. We demonstrate the effectiveness of the dynamic filter network on the tasks of video and stereo prediction, and reach state-of-the-art performance on the moving MNIST dataset with a much smaller model. By visualizing the learned filters, we illustrate that the network has picked up flow information by only looking at unlabelled training data. This suggests that the network can be used to pretrain networks for various supervised tasks in an unsupervised way, like optical flow and depth estimation.
We propose a function-based temporal pooling method that captures the latent structure of the video sequence data - e.g. how frame-level features evolve over time in a video. We show how the parameters of a function that has been fit to the video data can serve as a robust new video representation. As a specific example, we learn a pooling function via ranking machines. By learning to rank the frame-level features of a video in chronological order, we obtain a new representation that captures the video-wide temporal dynamics of a video, suitable for action recognition. Other than ranking functions, we explore different parametric models that could also explain the temporal changes in videos. The proposed functional pooling methods, and rank pooling in particular, is easy to interpret and implement, fast to compute and effective in recognizing a wide variety of actions. We evaluate our method on various benchmarks for generic action, fine-grained action and gesture recognition. Results show that rank pooling brings an absolute improvement of 7-10 average pooling baseline. At the same time, rank pooling is compatible with and complementary to several appearance and local motion based methods and features, such as improved trajectories and deep learning features.
In this paper we present a hierarchical method to discover mid-level elements with the objective of modeling visual compatibility between objects. At the base-level, our method identifies patterns of CNN activations with the aim of modeling different variations/styles in which objects of the classes of interest may occur. At the top-level, the proposed method discovers patterns of co-occurring activations of base-level elements that define visual compatibility between pairs of object classes. Experiments on the massive Amazon dataset show the strength of our method at describing object classes and the characteristics that drive the compatibility between them.