Exploiting Target Data to Learn Deep Convolutional Networks for Scene-Adapted Human Detection

Abstract

The difference between sample distributions of public data sets and specific scenes can be very significant. As a result, the deployment of generic human detectors in real-world scenes most often leads to sub-optimal detection performance. To avoid the labor-intensive task of manual annotations, we propose a semi-supervised approach for training deep convolutional networks on partially labeled data. To exploit a large amount of unlabeled target data, the knowledge learnt from public data sets is transferred to new model training by adapting an auxiliary detector to the target scene. We hypothesize that the components of the auxiliary detector capture essential human characteristics useful for constructing a scene-adapted detector. A selective ensemble algorithm is proposed to select a subset of the components relevant to the target scene for recombination. The resulting model is applied for collecting high-confidence samples from unlabeled target data. Furthermore, a deep convolutional network is trained by progressively labeling and selecting new training samples in a self-paced way. The detailed experimental evaluation verifies the effectiveness and superiority of the proposed approach in scene-specific human detection.

Self-Enhanced R-CNNs for Human Detection With Semi-Supervised Assumptions

“…Among those methods using unlabeled data but without any annotation, 'SMC Faster R-CNN' [45] performs the best on both datasets. When 5% of the training images are annotated, better detection results can be obtained by the semi-supervised methods: 'Variant SemiBoost' [14], 'Temporal Ensembling' [58] and 'Self-paced CNN' [15]. In this case, the proposed approach is able to outperform all the three semi-supervised methods.…”

“…In the case of limited labeled data, Wu et al [14] utilized the similarity between labeled data and unlabeled data to improve the training of boosted forests. In another work [15], a self-paced learning paradigm was adopted to progressively train a CNN by incrementally including more pseudo-labeled data.…”

“…Exploiting a large amount of unlabeled data is helpful for collecting adequate number of object instances, which is in turn useful for improving detection performance. For human detection, previous methods [13]- [15] mainly adopt a two-stage strategy: apply a pre-trained detector to unlabeled data for new training sample collection, and re-train the detector on the extended training data. In contrast to these methods, we propose a single-stage approach to feed both labeled data and unlabeled data directly to a Region-based Convolutional Neural Network (R-CNN) for enhancing detector training, such that the training process is simplified and the detection performance can be improved significantly.…”

Algorithm for Target Detection in Smart City Combined with Depth Learning and Feature Extraction

“…In recent years, more and more researchers have begun to conduct in-depth research on deep learning algorithms in target detection [1,2]. Deep learning avoids the drawbacks of the traditional method of manually extracting features, because of the characteristics that its deep structure can effectively learn from large amounts of data [3,4]. Currently, based on the target detection algorithm, the depth study of literature is not much.…”

A scale-aware YOLO model for pedestrian detection

“…However, this type of approach has a high detection error rate. Recently, convolutional neural networks (CNNs) have shown significant performance in a range of different applications, with pedestrian detection being one of the key areas where CNNs clearly outperforms traditional approaches [3,4,5,6,7]. For example, in [6], an end-to-end CNN architecture is employed to generate pedestrian bounding boxes via multiple layers in an image, and a classifier performs classification on bounding boxes.…”