Predicting Failures of Vision Systems

By Peng Zhang, Jiuling Wang, Ali Farhadi, Martial Hebert, Devi Parikh. 2014

135 citations recorded by the source corpus.

Computer vision systems today fail frequently. They also fail abruptly without warning or explanation. Alleviating the former has been the primary focus of the community. In this work, we hope to draw the community's attention to the latter, which is arguably equally problematic for real applications. We promote two metrics to evaluate failure prediction. We show that a surprisingly straightforward and general approach, that we call ALERT, can predict the likely accuracy (or failure) of a variety of computer vision systems - semantic segmentation, vanishing point and camera parameter estimation, and image memorability prediction - on individual input images. We also explore attribute prediction, where classifiers are typically meant to generalize to new unseen categories. We show that ALERT can be useful in predicting failures of this transfer. Finally, we leverage ALERT to improve the performance of a downstream application of attribute prediction: zero-shot learning. We show that ALERT can outperform several strong baselines for zero-shot learning on four datasets.

Subfields: Artificial Intelligence, 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 Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Towards Open Set Deep Networks (2015, methodology): “, we want to add a meta-recognition algorithm [22, 32] to analyze scores and recognize when deep networks are likely incorrect in their assessment.”
Examining CNN representations with respect to Dataset Bias (2017, methodology): “However, we believe that compared to (Bansal, Farhadi, and Parikh 2014; Zhang et al. 2014), it is of larger value to explore evidence of failure cases from mid-layer representations of a CNN.”
BranchyNet: Fast inference via early exiting from deep neural networks (2016, methodology): “Additionally, it may be possible to use a Meta-Recognition algorithm [19, 26] to estimate the characteristics of unseen test samples and adjust T automatically in order to maintain a specified runtime or accuracy goal.”
Question Relevance in VQA: Identifying Non-Visual And False-Premise Questions (2016, background): “Such failure prediction systems have been explored in vision (Zhang et al., 2014; Devarakota et al., 2007) and speech (Zhao et al., 2012; Sarma and Palmer, 2004; Choularton, 2009; Voll et al., 2008).”
Do explanations make VQA models more predictable to a human? (2018, methodology): “While Bansal et al. (2014) and Zhang et al. (2014) predict failures of a model using simpler statistical models, we explicitly train a person to do this.”
Relaxed Softmax: Efficient Confidence Auto-Calibration for Safe Pedestrian Detection (2018, background): “Several authors studied the reliability of classical logistic regression [1] and SVM [13, 29, 7] classiﬁers.”
3D ShapeNets: A deep representation for volumetric shapes (2014, background): “ave to resolve to another view to gather more observation for recognizing an object. This situation is even more common for computer vision. Automatic object recognition systems today fail frequently [28], and we desire a robust system that can recover from errors automatically. In particular, as sh”
It Takes Two to Tango: Towards Theory of AI's Mind (2017, background): “Their ability to estimate Vicki’s success or failure in a scenario is a measure of how well they understand its strengths and weaknesses.”
Are Elephants Bigger than Butterflies? Reasoning about Sizes of Objects (2016, unknown): “Figure 3c shows the precision of the models vs. declaration rate (Zhang et al. 2014).”
Semantic bottleneck for computer vision tasks (2018, methodology): “n this work, we propose to evaluate the explainability power of the semantic bottleneck by measuring its capacity to detect failure of the prediction function, either through an automated detector as [6], or through human judgment. Our proposal to generate the surrogate semantic representation is to”
Best of both worlds: Human-machine collaboration for object annotation (2015, unknown): “, [65] may be useful).”
Mirror, mirror on the wall, tell me, is the error small? (2015, methodology): “As a method that estimates the quality of the output of a vision system, our method is related to works like the meta-recognition [16], face recognition score analysis [22] and the recent failure alert [29] for failure prediction.”
Introspective perception: Learning to predict failures in vision systems (2016, methodology): “[20] learn a SVM model to predict failures from visual input.”
Improving Confidence Estimates for Unfamiliar Examples (2018, unknown): “Other methods aim to reduce conﬁdent errors by detecting failure [8, 10, 55, 51, 44], for example by looking at how close samples are to decision boundaries or estimating whether a test sample comes from the same distribution as training [5, 25, 28, 48, 22].”
Towards Accountable AI: Hybrid Human-Machine Analyses for Characterizing System Failure (2018, methodology): “Previous work in error detection (Zhang et al. 2014; Bansal, Farhadi, and Parikh 2014) (Nushi et al. 2017).”
3D ShapeNets for 2.5D Object Recognition and Next-Best-View Prediction (2014, background): “Automatic object recognition systems today fail frequently [27], and we desire a robust system that can recover from errors automatically.”
Association for Computational Linguistics " Why Should I Trust You? " Explaining the Predictions of Any Classifier (2016, unknown): “Some recent work aims to anticipate failures in machine learning, specifically for vision tasks [3, 29].”