Taking a deeper look at pedestrians

By Jan Hendrik Hosang, Mohamed Omran, Rodrigo Benenson, Bernt Schiele. Computer Vision and Pattern Recognition, 2015

352 citations recorded by the source corpus.

In this paper we study the use of convolutional neural networks (convnets) for the task of pedestrian detection. Despite their recent diverse successes, convnets historically underperform compared to other pedestrian detectors. We deliberately omit explicitly modelling the problem into the network (e.g. parts or occlusion modelling) and show that we can reach competitive performance without bells and whistles. In a wide range of experiments we analyse small and big convnets, their architectural choices, parameters, and the influence of different training data, including pretraining on surrogate tasks. We present the best convnet detectors on the Caltech and KITTI dataset. On Caltech our convnets reach top performance both for the Caltech1x and Caltech10x training setup. Using additional data at training time our strongest convnet model is competitive even to detectors that use additional data (optical flow) at test time.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/3054aa4f655c279cc79223970f80b11c67231b0b

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Deep Learning for Generic Object Detection: A Survey (2018, background): “Since our focus is on generic object detection, the extensive work on DCNNs for specific object detection, such as face detection [154, 306, 116], pedestrian detection [307, 109], vehicle detection [322] and traffic sign detection [329] will not be considered.”
UA-DETRAC: A new benchmark and protocol for multi-object detection and tracking (2015, methodology): “The minimum size of the detected object is set to 25 × 25 pixels 7 Since the selective search method is less e ﬀ ective in generating accurate region proposals of vehicles, we use the outputs of ACF as proposals (Hosang et al., 2015) for the R-CNN method.”
Pedestrian Detection: The Elephant In The Room (2020, methodology): “The same trend is also true for the pedestrian detection [2,17,7].”
CityPersons: A Diverse Dataset for Pedestrian Detection (2017, background): “Initial attempts to apply convnets for pedestrian detection, used existing detectors (mainly decision forests over hand-crafted features [2, 33]) outputs and re-scored them with a convnet classifier (plus bounding box regression) [16, 28, 1, 27].”
Object Detection in 20 Years: A Survey (2019, background): “Pedestrian detection is one of the first computer vision task that applies deep learning [312].”
Learning Efficient Single-Stage Pedestrian Detectors by Asymptotic Localization Fitting (2018, methodology): “[19] firstly utilizes the SCF detector [2] to generate proposals which are then fed into a RCNN-style network.”
Learning Complexity-Aware Cascades for Deep Pedestrian Detection (2015, methodology): “The “Proposal” columns report to the use of the CompACT cascade as a proposal mechanism [12, 15] for the CNN.”
UnitBox: An Advanced Object Detection Network (2016, background): “Currently, most of the CNN-based object detection meth-ods [2] [4] [8] could be summarized as a three-step pipeline: ﬁrstly, region proposals are extracted as object candidates from a given image.”
Deep Learning Strong Parts for Pedestrian Detection (2015, methodology): “[14] demonstrated the effectiveness of the R-CNN pipeline [12] in pedestrian detection and achieved top performance on Caltech [8] and KITTI[11].”
How Far are We from Solving Pedestrian Detection? (2016, methodology): “1) The AlexNet from [14], and 2) The VGG16 model from [11].”
Joint Detection and Identification Feature Learning for Person Search (2016, background): “Various factors, including CNN model structures, training data, and different training strategies are studied empirically in [15].”
Illumination-aware Faster R-CNN for Robust Multispectral Pedestrian Detection (2018, background): “Most of initial attempts [5, 6, 17, 18, 19] to apply convnets for pedestrian detection task adopted a two-stage pipeline in an R-CNN [20] style.”
Bi-box Regression for Pedestrian Detection and Occlusion Estimation (2018, methodology): “Many efforts have been made to improve its performance in recent years [3, 8, 17, 6, 33, 40, 5, 39, 41, 37, 34, 4].”
End-to-End Deep Learning for Person Search (2016, background): “Many person reidentification datasets have emerged in recent years, including VIPeR [9], ETHZ [29], i-LIDS [39], PRID2011 [13], RGB-D [2], GRID [22], CUHK01 [17], CUHK02 [16], and Multi-Camera Surveillance Database [3].”
Multispectral Deep Neural Networks for Pedestrian Detection (2016, methodology): “Along this research stream, many DNNs architectures have been investigated, including Ci-farNet [16], AlexNet [20], and VGG-16 [34].”
Occluded Pedestrian Detection Through Guided Attention in CNNs (2018, background): “Early works [12, 30] applying convnets for pedestrian detection are based on the RCNN structure [9], which relies on high-quality external proposals to achieve good performance.”
Scale-Aware Fast R-CNN for Pedestrian Detection (2015, methodology): “For fair comparison with the state-of-the-arts [17], [35], we also use the object proposals from ACF detector as inputs.”