Italy

This work investigates questions related to learning features in convolutional neural networks (CNN). Empirical findings across multiple architectures such as VGG, ResNet, Inception and MobileNet indicate that weights near the center of a filter are larger than weights on the outside. Current regularization schemes violate this principle. Thus, we introduce Locality-promoting Regularization, which yields accuracy gains across multiple architectures and datasets. We also show theoretically that the empirical finding could be explained by maximizing feature cohesion under the assumption of spatial locality.

ICPR 2020

LOCO-Reg: Locolaty promoting regularization

locality

representation learning

ICPR 2020 is the premier international conference on Pattern Recognition, gathering leading researchers from around the world to present their latest achievements in Pattern Analysis and Recognition, applied Artificial Intelligence and Machine Learning. The topic of the year is " Putting Artificial intelligence to work on Patterns". 
For the first time, ICPR has employed a two-round review process similar to journal submissions, with new papers allowed to be submitted in either the first or second round and papers submitted in the first round and not accepted allowed to be revised and re-submitted for second round review. 

In this edition of the **ICPR 2020** Conference, there are 1407 Oral/Poster presentations organized in five Tracks: 
- Artificial Intelligence, Machine Learning for Pattern Analysis; 
- Biometrics, Human Analysis and Behavior Understanding; Computer Vision, Robotics and Intelligent Systems; 
- Document and Media Analysis; and 
- Image and Signal Processing. 

The **ICPR 2020** Main Conference program also includes 7 Keynotes from prominent world-renowned scientists, 8 Tutorials, 28 Demos, 8 Challenges, 2 Panels, as well as an Industry/Sponsor Exhibition. In addition, 41 Workshops on a range of important current and emerging research topics complete the ICPR 2020 program. 


poster

Differently from the regular classification task, in ordinal classification there is an order in the classes. As a consequence, not all classification errors matter the same: a predicted class close to the ground truth one is better than predicting a farther away class. To account for this, most previous works employ loss functions based on the absolute difference between the predicted and ground truth class labels. We argue that there are many cases in ordinal classification where label values are arbitrary (for instance 1...C being C the number of classes) and thus such loss functions may not be the best choice. We instead propose a network architecture that produces not a single class prediction but an ordered vector, or ranking, of all the possible classes from most to least likely. This is thanks to a loss function that compares ground truth and predicted rankings of these class labels, not the labels themselves. Another advantage of this new formulation is that we can enforce consistency in the predictions, namely, predicted rankings come from some unimodal vector of scores with mode at the ground truth class. We compare with the state of the art ordinal classification methods, showing that ours attains equal or better performance, as measured by common ordinal classification metrics, on three benchmark datasets. Furthermore, it is also suitable for a new task on image aesthetics assessment, i.e. most voted score prediction. Finally, we also apply it to building damage assessment from satellite images, providing an analysis of its performance depending on the degree of imbalance of the dataset.

Rank-Based Ordinal Classification

We address the problem of detecting anomalies with respect to structured patterns. To this end, we conceive a novel anomaly detection method called PIF, that combines
the advantages of adaptive isolation methods with the flexibility of preference embedding. Specifically, we propose to embed the data in a high dimensional space where an efficient tree-based method, PI-FOREST , is employed to compute an anomaly score. Experiments on synthetic and real datasets demonstrate that
PIF favorably compares with state-of-the-art anomaly detection techniques, and confirm that PI-FOREST is better at measuring arbitrary distances and isolate points in the preference space.

PIF: Anomaly detection via preference embedding

Performance of fingerprint recognition algorithms substantially rely on fine features extracted from fingerprints. Apart from minutiae and ridge patterns, pore features have proven to be usable for fingerprint recognition. Although features from minutiae and ridge patterns are quite attainable from low-resolution images, using pore features is practical only if the fingerprint image is of high resolution which necessitates a model that enhances the image quality of the conventional
500 ppi legacy fingerprints preserving the fine details. To find a solution for recovering pore information from low-resolution fingerprints, we adopt a joint learning-based approach that combines both super-resolution and pore detection networks. Our modified single image Super-Resolution Generative Adversarial
Network (SRGAN) framework helps to reliably reconstruct high-resolution fingerprint samples from low-resolution ones assisting the pore detection network to identify pores with a high accuracy. The network jointly learns a distinctive feature representation from a real low-resolution fingerprint sample and successfully synthesizes a high-resolution sample from it. To add discriminative information and uniqueness for all the subjects, we have integrated features extracted from a deep fingerprint verifier with the SRGAN quality discriminator. We also add ridge reconstruction loss, utilizing ridge patterns to make the best use
of extracted features. Our proposed method solves the recognition problem by improving the quality of fingerprint images. High recognition accuracy of the synthesized samples that is close to the accuracy achieved using the original high-resolution images validate the effectiveness of our proposed model.

Super-resolution Guided Pore Detection for Fingerprint Recognition


Simultaneous localization and mapping (SLAM) algorithms aimed for autonomous vehicles (AVs) are required to utilize sensor redundancies specific to AVs and enable accurate, fast and repeatable estimations of pose and path trajectories. In this work, we present a combination of three SLAM algorithms that utilize a different subset of available sensors such as inertial measurement unit (IMU), a gray-scale mono-camera, and a Lidar. Also, we propose a novel acceleration-based gravity direction initialization (AGI) method for the visual-inertial SLAM algorithm. We analyze the SLAM algorithms and initialization methods for pose estimation accuracy, speed of convergence and repeatability on the KITTI odometry sequences. The proposed VI-SLAM with AGI method achieves relative pose errors less than 2\%, convergence in half a minute or less and convergence time variability less than 3s, which makes it preferable for AVs.

AV-SLAM: Autonomous Vehicle Simultaneous Localization and Mapping with Gravity Direction Initialization

Graphs are non-euclidean structures that can represent many relational data efficiently. Many studies have proposed the convolution and the pooling operators on the non-euclidean domain. The graph convolution operators have shown astounding performance on various tasks such as node representation and classification. For graph classification, different pooling techniques are introduced, but none of them has considered both neighborhood of the node and the long-range dependencies of the node. In this paper, we propose a novel graph pooling layer R2POOL, which balances the structure information around the node as well as the dependencies with far away nodes. Further, we propose a new training strategy to learn coarse to fine representations. We add supervision at only intermediate levels to generate predictions using only intermediate-level features. For this, we propose the concept of an alignment score. Moreover, each layer's prediction is controlled by our proposed branch training strategy. This complete training helps in learning dominant class features at each layer for representing graphs. We call the combined model by R2MAN. Experiments show that R2MAN has the potential to improve the performance of graph classification on various datasets.


Region and Relations Based Multi Attention Network for Graph Classification


we propose to associate the 3D human pose, the 2D human pose projection and the 2D image appearance through a new orthographic projection based linear regression module. Unlike existing reprojection based approaches, our orthographic projection and regression do not suffer from small angle problems, which usually lead to overfitting in the depth dimension. The proposed method shows state-of-the-art performance on the Human3.6M dataset and generalizes well to in-the-wild images.

Orthographic Projection Linear Regression for Single Image 3D Human Pose Estimation

Face recognition in unconstrained environments is challenging due to variations in illumination, quality of sensing, motion blur and etc. An individual’s face appearance can vary drastically under different conditions creating a gap between train (source) and varying test (target) data. The domain gap could cause decreased performance levels in direct knowledge transfer from source to target. Despite fine-tuning with domain specific data could be an effective solution, collecting and annotating data for all domains is extremely expensive. To this end, we propose a self-supervised domain learning (SSDL) scheme that trains on triplets mined from unlabelled data. A key factor in effective discriminative learning, is selecting informative triplets. Building on most confident predictions, we follow an “easy-to-hard” scheme of alternate triplet mining and self-learning. Comprehensive experiments on four different benchmarks show that SSDL generalizes well on different domains.

SSDL: Self-Supervised Domain Learning for Improved Face Recognition

Detection of moving objects is a critical component of many computer vision tasks. Recently, deep learning architectures have been developed for supervised learning based moving object change detection. Some top performing architectures, like FgSegNet are single frame spatial appearance cue-based detection and tend to overfit to the training videos. We propose a novel compact multi-cue autoencoder deep architecture, Motion U-Net (MU-Net) for robust moving object detection that generalizes much better than FgSegNet and requires nearly 30 times fewer weight parameters. Motion and change cues are estimated using a multi-modal background subtraction module combined with flux tensor motion estimation. MU-Net was trained and evaluated on the CDnet-2014 change detection challenge video sequences and had an overall F-measure of 0.9369. We used the unseen SBI-2015 video dataset to assess generalization capacity where MU-Net had an F-measure of 0.7625 while FgSegNet v2 was 0.3519, less than half the MU-Net accuracy.

Motion U-Net: Multi-cue Encoder-Decoder Network for Motion Segmentation

In this paper, we propose a deep learning model based on deep multi-view representation learning, to address the video object segmentation task. The proposed model emphasizes the importance of the inherent correlation between video frames and incorporates a multi-view representation learning based on deep canonically correlated autoencoders. The multi-view representation learning in our model provides an efficient mechanism for capturing inherent correlations by jointly extracting useful features and learning better representation into a joint feature space, i.e., shared representation. To increase the training data and the learning capacity, we train the proposed model with pairs of video frames, i.e., Fa and Fb. During the segmentation phase, the deep canonically correlated autoencoders model encodes useful features by processing multiple reference frames together, which is used to detect the frequently reappearing. Our model enhances the state-of-the-art deep learning-based methods that mainly focus on learning discriminative foreground representations over appearance and motion. Experimental results over two large benchmarks demonstrate the ability of the proposed method to outperform competitive approaches and to reach good performances, in terms of semantic segmentation.

Video semantic segmentation using deep multi-view representation learning

In this paper we present a sequential approach to jointly retrieve camera auto-calibration, camera pose and the 3D reconstruction of a non-rigid object from an uncalibrated RGB image sequence, without assuming any prior information about the shape structure, nor the need for a calibration pattern, nor the use of training data at all. To this end, we propose a Bayesian filtering approach based on a sum-of-Gaussians filter composed of a bank of extended Kalman filters (EKF). For every EKF, we make use of dynamic models to estimate its state vector, which later will be Gaussianly combined to achieve a global solution. To deal with deformable objects, we incorporate a mechanical model solved by using the finite element method. Thanks to these ingredients, the resulting method is both efficient and robust to several artifacts such as missing and noisy observations as well as sudden camera motions, while being available for a wide variety of objects and materials, including isometric and elastic shape deformations. Experimental validation is proposed in real experiments, showing its strengths with respect to competing approaches.

Downloads

Next from ICPR 2020

Rank-Based Ordinal Classification

Similar lecture

JECL: Joint Embedding and Cluster Learning for Image-Text Pairs