Finding Tiny Faces

Abstract

Figure 1: We describe a detector that can find around 800 faces out of the reportedly 1000 present, by making use of novel characterizations of scale, resolution, and context to find small objects. Detector confidence is given by the colorbar on the right: can you confidently identify errors? AbstractThough tremendous strides have been made in object recognition, one of the remaining open challenges is detecting small objects. We explore three aspects of the problem in the context of finding small faces: the role of scale invariance, image resolution, and contextual reasoning. While most recognition approaches aim to be scale-invariant, the cues for recognizing a 3px tall face are fundamentally different than those for recognizing a 300px tall face. We take a different approach and train separate detectors for different scales. To maintain efficiency, detectors are trained in a multi-task fashion: they make use of features extracted from multiple layers of single (deep) feature hierarchy. While training detectors for large objects is straightforward, the crucial challenge remains training detectors for small objects. We show that context is crucial, and define templates that make use of massively-large receptive fields (where 99% of the template extends beyond the object of interest). Finally, we explore the role of scale in pre-trained deep networks, providing ways to extrapolate networks tuned for limited scales to rather extreme ranges. We demonstrate state-of-the-art results on massively-benchmarked face datasets (FDDB and WIDER FACE). In particular, when compared to prior art on WIDER FACE, our results reduce error by a factor of 2 (our models produce an AP of 82% while prior art ranges from 29-64%).

A Comprehensive Performance Evaluation of Deformable Face Tracking “In-the-Wild”

“…Some of the top-performing systems consist of commercial software, thus we did use the deep methods of Hu and Ramanan (2016), that are available as open source with the method of Hu and Ramanan (2016) reporting the latest best performance in FDDB. Additionally, we employ the top performing SVM-based method for learning rigid templates (King 2015), the best weakly and strongly supervised DPM implementations of Mathias et al (2014) and Zhu and Ramanan (2012), along with the best performing exemplarbased technique of Kumar et al (2015) .…”

Contextual Region-Based Convolutional Neural Network with Multilayer Fusion for SAR Ship Detection

“…More than 85% of ships have an area smaller than 8000 m 2 , that is, around 80 pixels on a SAR image, which is less than the object size of the ImageNet dataset (more than 80% of objects have sizes between 40 and 140 pixels) [33]. Additionally, the ships which offer AIS information have an average length of 168.3 m. Furthermore, the average area is around 51 pixels which is far less than the area that is able to cause a response on the last convolutional layer of VGG16.…”

“…For instance, an object located on land is highly unlikely to be considered a ship, while an object with bright intensity in the ocean area is prone to be affirmed as a positive object. In order to mimic the visual effect of a human being in a computer vision field, context information is always added into the deep neural network to recognize the small-sized objects [27,29,33].…”

Comparing apples and oranges: Off‐road pedestrian detection on the National Robotics Engineering Center agricultural person‐detection dataset

“…We follow the approach in Ref. and use “oversized” templates, whose spatial support includes background pixels surrounding the object of interest, shown as contextualized templates in Figure . It turns out that including massive amounts of surrounding area (such that 99% of the template includes the background), which may capture additional contextual cues, such as shadows from a ground plane, is helpful for finding small objects.…”

“…We call the method Multiscale Foveal Context (MFC) in our results and refer the reader to Ref. for more quantitative analysis, such as how different ways of encoding context affects performance.…”