Do We Need More Training Data or Better Models for Object Detection?

By Xiangxin Zhu, Carl Vondrick, Deva Ramanan, Charless C. Fowlkes. 2012

191 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Evaluation & Benchmarks

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

In Evaluation & Benchmarks

Citations in the corpus

Fast R-CNN (2015, background): “In [23], Zhu et al . studied this issue in depth for DPMs [7] and found that they saturate after only a few hundred to thousand training examples.”
Object Detection Networks on Convolutional Feature Maps (2015, methodology): “PM+CNN methods [24 ,11 29] and R-CNN/SPPnet. The DPM model is a less powerful classiﬁer than a multi-layer network, whether applied to deep convolutional features or shallow HOG features. The work in [34] shows that a DPM classiﬁer has diminishing improvement when given more training data. To method”
Microsoft COCO: Common Objects in Context (2014, background): “Consistent with past observations [46], we ﬁnd that including difﬁcult (non-iconic) images during training may not always help.”
Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation (2013, background): “Additionally, [42] shows that even when more data is available, DPM does not easily benefit from it.”
A Constructive Prediction of the Generalization Error Across Scales (2019, background): “Data scaling: It has long been recognized that more data improves performance, and various studies report such trends in both computer vision (e.g., Zhu et al., 2012; Sun et al., 2017) and language processing tasks (e.g., Banko & Brill, 2001; Talmor & Berant, 2019).”
HOGgles: Visualizing Object Detection Features (2013, background): “[24] try to determine whether we have reached Bayes risk for HOG.”
From generic to specific deep representations for visual recognition (2014, background): “Zhu et al.[48] suggest that increasing data is less effective than increasing the complexity of models or richness of representation and the former is prone to early performance saturation.”
Are all training examples equally valuable? (2013, background): “More recently, Zhu et al. [22] found that off-the-shelf detectors trained with a “clean” subset of the training set may achieve higher performance than the full set.”
Random forest as a generic framework for predictive modeling of spatial and spatio-temporal variables (2018, background): “It does not have an718 explicit (geo)statistical model as a starting point, so that no mathematical derivations are possible at all.719 Also, just adding more and more data to the system does not necessarily mean that the predictions will720 automatically become better (Zhu et al., 2012).”
Understanding Objects in Detail with Fine-Grained Attributes (2014, unknown): “3.2 also shows that, contrary to what has been observed in other object detection tasks in the literature [45], detection of object parts can be improved by adding a significant number of part templates to a state-of-the-art detector.”
Deep Learning and Convolutional Neural Networks for Medical Image Computing (2017, methodology): “The trade-off between using better learning models and using more training data [49] should be carefully considered when searching for an optimal solution to any CADe problem (e.”
Human Pose Estimation Using Body Parts Dependent Joint Regressors (2013, methodology): “While this approach outperforms classical pictorial structure models for human pose estimation, it has been shown in [41] that the used templates, which are scanningwindow templates trained with linear SVMs on HOG features [7], are very sensitive to noise and limit the performance.”
Body Parts Dependent Joint Regressors for Human Pose Estimation in Still Images (2014, unknown): “Index Terms—Human pose estimation, fashion, random forest, regression, classification F”
Regularized Max Pooling for Image Categorization (2014, background): “If the locations of parts are unknown, the performance significantly drops [32].”
Detecting Avocados to Zucchinis: What Have We Done, and Where Are We Going? (2013, background): “[25] analyzed the relative impact of adding more training data versus building better detection models.”
Training Deformable Part Models with Decorrelated Features (2013, background): “Surprisingly, no previous work has studied how DPM detection accuracy varies as a function of the number of negative training examples ( cf. [27]).”
Learning Everything about Anything: Webly-Supervised Visual Concept Learning (2014, background): “These annotations can only tackle simple appearance variations of an object [51].”