Improving Semantic Segmentation via Efficient Self-Training.

Starting from the seminal work of Fully Convolutional Networks (FCN), there has been significant progress on semantic segmentation. However, deep learning models often require large amounts of pixelwise annotations to train accurate and robust models. Given the prohibitively expensive annotation cost of segmentation masks, we introduce a self-training framework in this paper to leverage pseudo labels generated from unlabeled data. In order to handle the data imbalance problem of semantic segmentation, we propose a centroid sampling strategy to uniformly select training samples from every class within each epoch. We also introduce a fast training schedule to alleviate the computational burden. This enables us to explore the usage of large amounts of pseudo labels. Our Centroid Sampling based Self-Training framework (CSST) achieves state-of-the-art results on Cityscapes and CamVid datasets. On PASCAL VOC 2012 test set, our models trained with the original train set even outperform the same models trained on the much bigger augmented train set. This indicates the effectiveness of CSST when there are fewer annotations. We also demonstrate promising few-shot generalization capability from Cityscapes to BDD100K and from Cityscapes to Mapillary datasets.

Machine Learning Based Autism Spectrum Disorder Detection from Videos.

Early diagnosis of Autism Spectrum Disorder (ASD) is crucial for best outcomes to interventions. In this paper, we present a machine learning (ML) approach to ASD diagnosis based on identifying specific behaviors from videos of infants of ages 6 through 36 months. The behaviors of interest include directed gaze towards faces or objects of interest, positive affect, and vocalization. The dataset consists of 2000 videos of 3-minute duration with these behaviors manually coded by expert raters. Moreover, the dataset has statistical features including duration and frequency of the above mentioned behaviors in the video collection as well as independent ASD diagnosis by clinicians. We tackle the ML problem in a two-stage approach. Firstly, we develop deep learning models for automatic identification of clinically relevant behaviors exhibited by infants in a one-on-one interaction setting with parents or expert clinicians. We report baseline results of behavior classification using two methods: (1) image based model (2) facial behavior features based model. We achieve 70% accuracy for smile, 68% accuracy for look face, 67% for look object and 53% accuracy for vocalization. Secondly, we focus on ASD diagnosis prediction by applying a feature selection process to identify the most significant statistical behavioral features and a over and under sampling process to mitigate the class imbalance, followed by developing a baseline ML classifier to achieve an accuracy of 82% for ASD diagnosis.

Predicting ASD Diagnosis in Children with Synthetic and Image-based Eye Gaze Data.

As early intervention is highly effective for young children with autism spectrum disorder (ASD), it is imperative to make accurate diagnosis as early as possible. ASD has often been associated with atypical visual attention and eye gaze data can be collected at a very early age. An automatic screening tool based on eye gaze data that could identify ASD risk offers the opportunity for intervention before the full set of symptoms is present. In this paper, we propose two machine learning methods, synthetic saccade approach and image based approach, to automatically classify ASD given children's eye gaze data collected from free-viewing tasks of natural images. The first approach uses a generative model of synthetic saccade patterns to represent the baseline scan-path from a typical non-ASD individual and combines it with the real scan-path as well as other auxiliary data as inputs to a deep learning classifier. The second approach adopts a more holistic image-based approach by feeding the input image and a sequence of fixation maps into a convolutional or recurrent neural network. Using a publicly-accessible collection of children's gaze data, our experiments indicate that the ASD prediction accuracy reaches 67.23% accuracy on the validation dataset and 62.13% accuracy on the test dataset.

PREDICTING AUTISM DIAGNOSIS USING IMAGE WITH FIXATIONS AND SYNTHETIC SACCADE PATTERNS.

Signs of autism spectrum disorder (ASD) emerge in the first year of life in many children, but diagnosis is typically made much later, at an average age of 4 years in the United States. Early intervention is highly effective for young children with ASD, but is typically reserved for children with a formal diagnosis, making accurate identification as early as possible imperative. A screening tool that could identify ASD risk during infancy offers the opportunity for intervention before the full set of symptoms is present. In this paper, we propose two machine learning methods, synthetic saccade approach and image based approach, to automatically classify ASD given the scanpath data from children on free viewing of natural images. The first approach uses a generative model of synthetic saccade patterns to represent the baseline scan-path from a typical non-ASD individual and combines it with the input scanpath as well as other auxiliary data as inputs to a deep learning classifier. The second approach adopts a more holistic image based approach by feeding the input image and a sequence of fixation maps into a state-of-the-art convolutional neural network. Our experiments indicate that we can get 65.41% accuracy on the validation dataset.