Poly-NL: Linear Complexity Non-local Layers With 3rd Order Polynomials

Spatial self-attention layers, in the form of Non-Local blocks, introduce long-range dependencies in Convolutional Neural Networks by computing pairwise similarities among all possible positions. Such pairwise functions underpin the effectiveness of non-local layers, but also determine a complexity that scales quadratically with respect to the input size both in space and time. This is a severely limiting factor that practically hinders the applicability of non-local blocks to even moderately sized inputs. Previous works focused on reducing the complexity by modifying the underlying matrix operations, however in this work we aim to retain full expressiveness of non-local layers while keeping complexity linear. We overcome the efficiency limitation of non-local blocks by framing them as special cases of 3rd order polynomial functions. This fact enables us to formulate novel fast Non-Local blocks, capable of reducing the complexity from quadratic to linear with no loss in performance, by replacing any direct computation of pairwise similarities with element-wise multiplications. The proposed method, which we dub as "Poly-NL", is competitive with state-of-the-art performance across image recognition, instance segmentation, and face detection tasks, while having considerably less computational overhead.

Localizing Polygonal Objects in Man-Made Environments

Xiaolu Sun

Object detection is a significant challenge in Computer Vision and has received a lot of attention in the field. One such challenge addressed in this thesis is the detection of polygonal objects, which are prevalent in man-made environments. Shape analysis is an important cue to detect these objects. We propose a contour-based object detection framework to deal with the related challenges, including how to efficiently detect polygonal shapes and how to exploit them for object detection. First, we propose an efficient component tree segmentation framework for stable region extraction and a multi-resolution line segment detection algorithm, which form the bases of our detection framework. Our component tree segmentation algorithm explores the optimal threshold for each branch of the component tree, and achieves a significant improvement over image thresholding segmentation, and comparable performance to more sophisticated methods but only at a fraction of computation time. Our line segment detector overcomes several inherent limitations of the Hough transform, and achieves a comparable performance to the state-of-the-art line segment detectors. However, our approach can better capture dominant structures and is more stable against low-quality imaging conditions. Second, we propose a global shape analysis measurement for simple polygon detection and use it to develop an approach for real-time landing site detection in unconstrained man-made environments. Since the task of detecting landing sites must be performed in a few seconds or less, existing methods are often limited to simple local intensity and edge variation cues. By contrast, we show how to efficiently take into account the potential sitesâ global shape, which is a critical cue in man-made scenes. Our method relies on component tree segmentation algorithm and a new shape regularity measure to look for polygonal regions in video sequences. In this way we enforce both temporal consistency and geometric regularity, resulting in reliable and consistent detections. Third, we propose a generic contour grouping based object detection approach by exploring promising cycles in a line fragment graph. Previous contour-based methods are limited to use additive scoring functions. In this thesis, we propose an approximate search approach that eliminates this restriction. Given a weighted line fragment graph, we prune its cycle space by removing cycles containing weak nodes or weak edges, until the upper bound of the cycle space is less than the threshold defined by the cyclomatic number. Object contours are then detected as maximally scoring elementary circuits in the pruned cycle space. Furthermore, we propose another more efficient algorithm, which reconstructs the graph by grouping the strongest edges iteratively until the number of the cycles reaches the upper bound. Our approximate search approaches can be used with any cycle scoring function. Moreover, unlike other contour grouping based approaches, our approach does not rely on a greedy strategy for finding multiple candidates and is capable of finding multiple candidates sharing common line fragments. We demonstrate that our approach significantly outperforms the state-of-the-art.

EPFL2015

Understanding Deep Neural Networks using Adversarial Attacks

Krishna Kanth Nakka

Deep Neural Networks (DNNs) have achieved great success in a wide range of applications, such as image recognition, object detection, and semantic segmentation. Even thoughthe discriminative power of DNNs is nowadays unquestionable, serious concerns have arised ever since DNNs have shown to be vulnerable to adversarial examples craftedby adding imperceptible perturbations to clean images. The implications of these malicious attacks are even more significant for DNNs deployed in real-world systems, e.g.,autonomous driving and biometric authentication. Consequently, an intriguing question that we aim to understand is the underlying behavior of DNNs to adversarial attacks.This thesis contributes to a better understanding of the mechanism of adversarial attacks on DNNs. Our main contributions are broadly in two directions: (1) we proposeinterpretable architectures first to understand the reasons for the success of adversarial attacks and then to improve the robustness of DNNs; (2) we design intuitive adversarialattacks to both mislead and use as a tool to expand our present understanding of DNNs' internal workings and their limitations. In the first direction, we introduce deep architectures that allow humans to interpret the reasoning process of DNNs prediction. Specifically, we incorporate Bag-of-visual-wordsrepresentations from the pre-deep learning era into DNNs using an attention scheme. We find key reasons for adversarial attack success and use these insights to propose anadversarial defense by maximally separating the latent features of discriminative regions while minimizing the contribution of non-discriminative regions in the final prediction.The second direction deals with the design of adversarial attacks to understand DNNs' limitations in a real-world environment. To begin with, we show that existing state-of-the-art semantic segmentation networks that achieve superior performance by exploiting the context are highly susceptible to indirect local attacks. Furthermore, we demonstrate the existence of universal directional perturbations that are quasi-independent of the input template but still successfully fool unknown siamese-based visual object trackers. We then identify that the mid-level filter banks across different backbones bear strong similarities and thus can be potential common ground for attack. We, therefore, learn a generator that disrupts mid-level features with high transferability across different target architectures, datasets, and tasks. In short, our attacks highlight critical vulnerabilities of DNNs, which make their deployment challenging in the real-world environment, even in the extreme case when the attacker is unaware of the target architecture or the targetdata used to train it.Furthermore, we go beyond fooling networks and demonstrate the usefulness of adversarial attacks for studying the internal disentangled representations in self-supervised 3D pose estimation networks. We observe that adversarial manipulation of appearance information in the input image alters the pose output, indicating that the pose code contains appearance information and disentanglement is far from complete. Besides the above contributions, an underlying theme that arises multiple times in this thesis is counteracting the adversarial attacks by detecting them.

EPFL2022

Learning Pose Invariant and Covariant Classifiers from Image Sequences

Mustafa Özuysal

Object tracking and detection over a wide range of viewpoints is a long-standing problem in Computer Vision. Despite significant advance in wide-baseline sparse interest point matching and development of robust dense feature models, it remains a largely open problem. Moreover, abundance of low cost mobile platforms and novel application areas, such as real-time Augmented Reality, constantly push the performance limits of existing methods. There is a need to modify and adapt these to meet more stringent speed and capacity requirements. In this thesis, we aim to overcome the difficulties due to the multi-view nature of the object detection task. We significantly improve upon existing statistical keypoint matching algorithms to perform fast and robust recognition of image patches independently of object pose. We demonstrate this on various 2D and 3D datasets. The statistical keypoint matching approaches require massive amounts of training data covering a wide range of viewpoints. We have developed a weakly supervised algorithm to greatly simplify their training for 3D objects. We also integrate this algorithm in a 3D tracking-by-detection system to perform real-time Augmented Reality. Finally, we extend the use of a large training set with smooth viewpoint variation to category-level object detection. We introduce a new dataset with continuous pose annotations which we use to train pose estimators for objects of a single category. By using these estimators' output to select pose specific classifiers, our framework can simultaneously localize objects in an image and recover their pose. These decoupled pose estimation and classification steps yield improved detection rates. Overall, we rely on image and video sequences to train classifiers that can either operate independently of the object pose or recover the pose parameters explicitly. We show that in both cases our approaches mitigate the effects of viewpoint changes and improve the recognition performance.

EPFL2010