On the Performance of GoogLeNet and AlexNet Applied to Sketches

By Pedro L. Ballester, Ricardo Matsumura de Araújo. 2016

339 citations recorded by the source corpus.

This work provides a study on how Convolutional Neural Networks, trained to identify objects primarily in photos, perform when applied to more abstract representations of the same objects. Our main goal is to better understand the generalization abilities of these networks and their learned inner representations. We show that both GoogLeNet and AlexNet networks are largely unable to recognize abstract sketches that are easily recognizable by humans. Moreover, we show that the measured efficacy vary considerably across different classes and we discuss possible reasons for this.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

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

In Computer Vision and Pattern Recognition

Citations in the corpus

ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness (2018, background): “CNNs can still recognise images with scrambled shapes (Gatys et al., 2017; Brendel & Bethge, 2019), but they have much more difﬁculties recognising objects with missing texture information (Ballester & de Ara´ujo, 2016; Yu et al., 2017).”
Interpreting Adversarially Trained Convolutional Neural Networks (2019, background): “Along this line, some recent works showed that standard CNNs trained on ImageNet make their predictions rely on the local textures rather than long-range dependencies encoded in the shape of objects (Geirhos et al., 2019; Brendel & Bethge, 2019; Ballester & de Ara´ujo, 2016).”
Knowledge Distillation via Instance Relationship Graph (2019, neutral): “Examples include from AlexNet [13, 11, 2] to ResNet [7, 27, 21] and DenseNet [10].”
Reducing Domain Gap via Style-Agnostic Networks (2019, background): “Some studies [12,25] have also shown that CNNs perform well when only local textures are given while global shape structures are missing, but work poorly in the reverse scenario [3,14].”
Reducing Domain Gap by Reducing Style Bias (2021, background): “Some studies have also shown that CNNs perform well when only local textures are given while global shape structures are missing [12, 6], but work poorly in the reverse scenario [3, 14].”
Deep Learning for Free-Hand Sketch: A Survey (2020, background): “Numerous deep models have now been proposed for free-hand sketch recognition [67], [87], [88], [89], [90], [91], [92], [93], [94], [95], [96], [97], [98].”
Inverting the Generator of a Generative Adversarial Network (2016, background): “, AlexNet) in the inversion process (see [3] for empirical evidence of bias even in visual networks).”
The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (2019, background): “Other studies have shown that CNNs are sensitive to a wide range of other image manipulations that have little effect on human judgments [32, 5, 23, 76, 37, 3, 44, 46, 1, 6].”