The Importance of Depth for Visual Tracking – A fundamental issue in all deep vision systems is to classify high-accuracy visual observations. Recent studies have found that deep neural networks outperform the state-of-the-art visual tracking methods in learning from images. This work investigates that deep neural networks also are useful in predicting the state of the visual system and the object category. The main motivation behind this work is to predict the category of the visual system. For this purpose I present a deep-feature learning method based on both feature predictions and a sparse and unsupervised feature representation of the system. Experiments show that the proposed method outperforms state-of-the-art visual tracking and object detection.
We develop a new method for predicting the performance of a deep neural network (DNN) trained on image classification over supervised learning. We first show that the prediction of the performance of a DNN using our method is indeed a good match for the problem. Then, we demonstrate the strength of our method, by testing on several commonly-used models including Deep CNN and ConvNets. Our results show that the proposed algorithm is very strong — we can predict the performance of over 70 CNN models.
On Unifying Information-based and Information-based Suggestive Word Extraction
Learning to Use Context Disparity in Learners for Retrieval and Learning
The Importance of Depth for Visual Tracking
Learning from the Hindsight Plan: On Learning from Exact Time-series Data
Multivariate Student’s Test for Interventional ErrorWe develop a new method for predicting the performance of a deep neural network (DNN) trained on image classification over supervised learning. We first show that the prediction of the performance of a DNN using our method is indeed a good match for the problem. Then, we demonstrate the strength of our method, by testing on several commonly-used models including Deep CNN and ConvNets. Our results show that the proposed algorithm is very strong — we can predict the performance of over 70 CNN models.