PhD Defense: Image and Video Understanding with Constrained Resources
The exponential growth of images and videos has encouraged the development of systems that can perform automated understanding of visual data both effectively and efficiently. Recent advances in computer vision tasks like image recognition and object detection have been driven by high-capacity deep neural networks, particularly Convolutional Neural Networks (CNNs) with hundreds of layers, trained in a supervised manner with clean and massive human annotations. However, this poses two significant challenges: (1) the increased depth in CNNs that leads to significant improvements over competitive benchmarks at the same time, limits their deployment in real-world scenarios due to high computational cost, especially for applications on mobile devices and delay-sensitive systems, where inputs need to be processed in real-time as they arrive; (2) the need to collect millions of human labeled samples for training prevents such approaches to scale, especially for fine-grained image understanding like semantic segmentation, where dense annotations are extremely expensive to obtain. To mitigate these issues, we focus on image and video understanding with constrained resources, in the forms of computational resources and annotation resources. In particular, we present approaches that (1) investigate dynamic computation frameworks which adaptively allocate computing resources on-the-fly given a novel image/video to manage the trade-off between accuracy and computational complexity; (2) derive robust representations with minimal human supervision through exploring context relationships or using shared information across domains.With this in mind, we first introduce BlockDrop, a conditional computation approach that learns to dynamically choose which layers of a deep network to execute during inference so as to best reduce total computation without degrading prediction accuracy. Exploiting the robustness of Residual Networks (ResNets) to layer dropping, our framework selects on-the-fly which residual blocks to evaluate for a given novel image. In particular, given a pretrained ResNet, we train a policy network in an associative reinforcement learning setting for the dual reward of utilizing a minimal number of blocks while preserving recognition accuracy.Next, we generalize the idea of conditional computation of images to videos by presenting AdaFrame, a framework that adaptively selects relevant frames on a per-input basis for fast video recognition. AdaFrame contains a Long Short-Term Memory network augmented with a global memory that provides context information for searching which frames to use over time. Trained with policy gradient methods, AdaFrame generates a prediction, determines which frame to observe next, and computes the utility, i.e., expected future rewards, of seeing more frames at each time step. At testing time, AdaFrame exploits predicted utilities to achieve adaptive lookahead inference such that the overall computational costs are reduced without incurring a decrease in accuracy.AdaFrame assumes access to all frames in videos, and hence can be only used in offline settings. To mitigate this issue, we introduce LiteEval, a simple yet effective coarse-to-fine framework for resource efficient video recognition, suitable for both online and offline scenarios. Exploiting decent yet computationally efficient features derived at a coarse scale with a lightweight CNN model, LiteEval dynamically decides on-the-fly whether to compute more powerful features for incoming video frames at a finer scale to obtain more details. This is achieved by a coarse LSTM and a fine LSTM operating cooperatively, as well as a conditional gating module to learn when to allocate more computation.To derive robust feature representations with limited annotation resources, we first explore the power of spatial context as a supervisory signal for learning visual representations. In particular, we present a spatial context network that is trained to predict a representation of one image patch from another image patch, within the same image, conditioned on their real-valued relative spatial offset. Once the spatial context network is trained, it can be further used for other tasks.In addition, we also propose to learn from synthetic data rendered by modern computer graphics tools, where ground-truth labels are readily available. We propose Dual Channel-wise Alignment Networks (DCAN), a simple yet effective approach to reduce domain shift at both pixel-level and feature-level, for unsupervised scene adaptation. DCAN leverages channel-wise feature alignment in both the image generator for synthesizing photo-realistic samples, appearing as if drawn from the target set, and the segmentation network, which simultaneously normalizes feature maps of source images.
Chair: Dr. Larry S. Davis Dean's rep: Dr. Rama Chellappa Members: Dr. David Jacobs Dr. Tom Goldstein
Dr. Abhinav Shrivastava