The recognition of mid-18th to mid-20th century piano scores presents segmentation challenges caused by touching and broken symbols produced by imprinting techniques and time degradation. We present a new notehead accidental dataset containing 2955 images from dense and damaged piano scores. We address this detection problem with very small training samples using a simple Spatial Transformer (ST)-based Convolutional Neural Network detector improved through bootstrapping and contextual information, and more powerful deep learning detectors (Faster R-CNN, R-FCN, and SSD) with transfer-learning on the COCO dataset. We trained all our detectors using 5 fold cross-validation and obtain 98.73% mean Average Precision (mAP) for an Intersection over Union (IoU) threshold of 0.75 with our best detector. Our ST-based detector obtains a slightly lower mAP of 94.81%, but runs 40 times faster, and uses 18 times less memory.

Optical Music Recognition (OMR) is the challenge of understanding the content of musical scores. Accurate detection of individual music objects is a critical step in processing musical documents because a failure at this stage corrupts any further processing. So far, all proposed methods were either limited to typeset music scores or were built to detect only a subset of the available classes of music symbols. In this work, we propose an end-to-end trainable object detector for music symbols that is capable of detecting almost the full vocabulary of modern music notation in handwritten music scores. By training deep convolutional neural networks on the recently released MUSCIMA++ dataset which has symbol-level annotations, we show that a machine learning approach can be used to accurately detect music objects with a mean average precision of over 80%.

This paper addresses the problem of analyzing human motion in an image sequence. This is of particular importance for video-surveillance applications. We have developed an original and efficient approach to track the apparent contours of a moving articulated structure, avoiding the use of 3D models. This method exploits spatio-temporal XT-slices from the image sequence volume XYT, where typical trajectory patterns can be associated with articulated motion such as human walking. We reconstruct these trajectories online by introducing an appropriate predictive model while correctly handling occlusion periods. This paradigm can lead to a simple trajectory recognition scheme. Experiments with real-world images depicting human walk or gesture are reported, and obtained results validate the proposed approach.

This paper is concerned with the problem of analysing articulated motion in image sequences, especially applied to case of the human motion. We have developed an approach to track the contours of a moving articulated structure, designed in a purely 2D+t form avoiding explicit and dedicated models. This method exploits spatio-temporal XT-slices of the image sequence volume XYT, where articulated motion, such as the gait of a walker, turns out to reveal typical trajectory providing relevant signatures. We propose a reconstruction algorithm of these significant trajectories, operating on-line and resorting to an original predictive model in order to deal with occultation periods. This paradigm provides the basis of trajectory classification and recognition schemes. Results are presented for real-world images depicting human movement of gestures and walking.

This paper is concerned with the problem of computing normal displacements along contours in image sequences. Our estimation is restricted to the perpendicular-to-the-edge velocity component, since the well-known "aperture problem" restricts any local estimation to this only component. We model moving edges as spatio-temporal surface patches in the image sequence space (x, y, t). A statistical regularization scheme based on Markov random fields allows us to get a homogeneous and relevant normal motion field along contours. It turns out that it can be implemented in an efficient way, mostly leading to convolution-like computations. Subpixel accuracy comes straightforwardly with this modeling, and is handled within the optimization stage itself, not as a post-processing step. Results are presented concerning synthetic experiments and real-world sequences.