Blocks That Shout: Distinctive Parts for Scene Classification

By Mayank Juneja, Andrea Vedaldi, C. V. Jawahar, Andrew Zisserman. 2013

428 citations recorded by the source corpus.

The automatic discovery of distinctive parts for an object or scene class is challenging since it requires simultaneously to learn the part appearance and also to identify the part occurrences in images. In this paper, we propose a simple, efficient, and effective method to do so. We address this problem by learning parts incrementally, starting from a single part occurrence with an Exemplar SVM. In this manner, additional part instances are discovered and aligned reliably before being considered as training examples. We also propose entropy-rank curves as a means of evaluating the distinctiveness of parts shareable between categories and use them to select useful parts out of a set of candidates. We apply the new representation to the task of scene categorisation on the MIT Scene 67 benchmark. We show that our method can learn parts which are significantly more informative and for a fraction of the cost, compared to previous part-learning methods such as Singh et al. [28]. We also show that a well constructed bag of words or Fisher vector model can substantially outperform the previous state-of-the-art classification performance on this data.

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

Unsupervised Visual Representation Learning by Context Prediction (2015, background): “Finally, our work is related to a line of research on discriminative patch mining [13, 47, 26, 34, 49, 11], which has emphasized weak supervision as a means of object discovery.”
CrossTransformers: spatially-aware few-shot transfer (2020, background): “Discriminative parts [5, 10, 22, 34, 42, 83] and visual words [74, 100] have a rich history, and have found applications in deformable-parts models [26], and classification [70, 100].”
CNN Features Off-the-Shelf: An Astounding Baseline for Recognition (2014, methodology): “The BoP approach [20] trains thousands of small part exemplar SVMs and expands a selection of those by retraining, it finally fuses them into a single representation and learns a new classifier on top of that.”
Analyzing the Performance of Multilayer Neural Networks for Object Recognition (2014, background): “, [12]), or by first finding a set of mid-level patches [10,20] and then encoding images in terms of them.”
Deep Image: Scaling up Image Recognition (2015, methodology): “fferent from outdoor scenes, it has large variability across different samples within each category. The work [39] created a dataset of 67 indoor scenes categories, which was used by many researchers [8, 15, 19, 26, 40]. Quattoni et al. propose a scene model which uses image prototypes to define a m”
Context as Supervisory Signal: Discovering Objects with Predictable Context (2014, background): “A number of recent works have argued that weak supervision can be an effective way to get more visually meaningful clusters [14,15,16,17,18,19,20,21,22,23].”
Multi-scale Orderless Pooling of Deep Convolutional Activation Features (2014, background): “In fact, several recent methods achieving state-of-the-art results on this dataset are based on the idea of finding such patches [38,37,36].”
Learning Visual Question Answering by Bootstrapping Hard Attention (2018, background): “One line of work on discriminative patches builds a representation by selecting some patches and ignoring others, which has proved useful for object detection and classification [56,57,58], and especially visualization [59].”
WILDCAT: Weakly Supervised Learning of Deep ConvNets for Image Classification, Pointwise Localization and Segmentation (2017, neutral): “The seminal DPM model [16] including several template regions for decision has been extensively studied [54, 46], optionally incorporating priors, e.g. sparsity or diversity, in order to learn sensible models [30, 59].”
On learning to localize objects with minimal supervision (2014, neutral): “Recent literature on unsupervised mid-level visual element discovery (Doersch et al., 2012; Singh et al., 2012; Endres et al., 2013; Juneja et al., 2013; Raptis et al., 2012) uses weak labels to discover visual elements that occur commonly in positive images but not in negative images.”
Patch to the Future: Unsupervised Visual Prediction (2014, background): “Our work builds upon the recent success of mid-level discriminative patch discovery [4, 5, 6, 16, 28] and proposes a prediction framework based on these mid-level elements.”
Exemplar based Deep Discriminative and Shareable Feature Learning for scene image classification (2015, background): “Comparing to current existing feature learning methods, they can generally get better local descriptors, and extra information (e.g. discriminative information) can be better expressed by manually inserting prior knowledge.”
Deep multiple instance learning for image classification and auto-annotation (2015, methodology): “ROI+Gist [30] 26.1 DPM [28] 30.4 Object Bank [22] 37.6 RBow [29] 37.9 BoP [14] 46.1 miSVM [23] 46.4 D-Parts [37] 51.4 IFV [14] 60.8 MLrep [8] 64.0 CNN-SVM [34] needs a feature representation of length over 200, 000, and MLrep [8] employs the very timeconsuming fine tuning on the dataset.”
Model recommendation: Generating object detectors from few samples (2015, background): “This would be true especially when considering the shared properties across instances and categories [30, 39, 12, 20].”
Contractive Rectifier Networks for Nonlinear Maximum Margin Classification (2015, neutral): “Improved Fisher Vectors [16] 60.”
Learning Discriminative Part Detectors for Image Classification and Cosegmentation (2013, neutral): “Part-based models have attracted much attention in computer vision recently [2, 4, 11, 15, 24, 32].”
Adaptive region pooling for object detection (2015, methodology): “However, these algorithms use rectangular patches to model object segments, which are less Figure 1.”