Places: An Image Database for Deep Scene Understanding

By Bolei Zhou, Aditya Khosla, Àgata Lapedriza, Antonio Torralba, Aude Oliva. 2016

492 citations recorded by the source corpus.

The rise of multi-million-item dataset initiatives has enabled data-hungry machine learning algorithms to reach near-human semantic classification at tasks such as object and scene recognition. Here we describe the Places Database, a repository of 10 million scene photographs, labeled with scene semantic categories and attributes, comprising a quasi-exhaustive list of the types of environments encountered in the world. Using state of the art Convolutional Neural Networks, we provide impressive baseline performances at scene classification. With its high-coverage and high-diversity of exemplars, the Places Database offers an ecosystem to guide future progress on currently intractable visual recognition problems.

Subfields: Computer Vision and Pattern Recognition, Media Technology

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

In Media Technology

Citations in the corpus

Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language (2022, methodology): “rank Places365 (Zhou et al. 2016) scene categories against the image, with the top n candidates (out of 365) inserted into a prefix: “Places: {place1}, {place2}, {place3}.”
Dual Path Networks (2017, methodology): “Key properties of the proposed DPNs are studied on the ImageNet-1k object classification dataset [17] and further verified on the Places365-Standard scene understanding dataset [24].”
Network Dissection: Quantifying Interpretability of Deep Visual Representations (2017, positive): “After training on Places365, the AlexNet-GAP-Wide obtains similar classiﬁcation accuracy on the validation set as the standard AlexNet ( 0 .”
Learning the Number of Neurons in Deep Networks (2016, methodology): “Furthermore, we conducted additional experiments on the character recognition dataset of [Jaderberg et al., 2014a Architectures: For ImageNet and Places2-401, our architectures are based on the VGG-B network (BNet) [Simonyan and Zisserman, 2014] and on DecomposeMe 8 (Dec 8 ) [Alvarez and Petersson, ”
Wider or Deeper: Revisiting the ResNet Model for Visual Recognition (2016, methodology): “ance. And it again shows why we should favour deeper models. 4) Model A2 performs worse than Model A on this dataset. We initialize it using weights from Model A, and tune it with the Places 365 data [39] for 45k iterations. This is reasonable since there are only object categories in this dataset, ”
Domain-Adaptive Deep Network Compression (2017, background): “One of the important factors in the success of deep learning for computer vision is the ease with which features, pre-trained on large datasets such as Imagenet [6, 31] and Places2 [37], can be transferred to other computer vision network (fc6, fc7).”
Detecting Visual Relationships with Deep Relational Networks (2017, background): “Thanks to the advances in deep learning, the past several years witness remarkable progress in several key tasks in computer vision, such as object recognition [2], scene classification [3], and attribute detection [4].”
Deep learning for visual understanding (2017, unknown): “We do not list the methods which employed additional information to improve classi cation, such as utilizing the part annotations for Oxford102, or using the large-scale scene-speci c Places2 dataset[250] for Indoor67/SUN397.”
Classification-Driven Dynamic Image Enhancement (2017, unknown): “thods and previous works on CUB. Dataset ConvNet FC-CNN Stat-CNN Dyn-CNN (ours) (ours) PasclVOC (mAP) AlexNet 76.9 [31] 78.48 83.06 PasclVOC (mAP) VGG-16 89.3 [28] 90.37 92.8 MIT (Acc.) AlexNet 56.79 [39] 58.24 62.9 MIT (Acc.) VGG-16 64.87 [39] 65.94 68.67 DTD (mAP) AlexNet 61.3 [5] 62.9 67.81 DTD (”
Image Inpainting using Block-wise Procedural Training with Annealed Adversarial Counterpart (2018, methodology): “We evaluate our inpainting method on Ima-geNet [22], COCO [18], Place2 [31] and CelebA [32].”
Knowledge Guided Disambiguation for Large-Scale Scene Classification With Multi-Resolution CNNs (2016, methodology): “The Places2 [19] is extended from the Places dataset, and is probably the largest scene recognition dataset currently3.”
Self-Supervised Visual Feature Learning With Deep Neural Networks: A Survey (2019, background): “Commonly used image datasets include ImageNet [14], Places [76], Places365 [77], SUNCG [79], MNIST [75], SVHN [80], CIFAR10 [74], STL-10 [78], PASCAL VOC [63], commonly used video datasets include YFCC100M [85], SceneNet RGB-D [84], Moment-in-Time [83], Kinetics [19], AudioSet [81], KITTI [82], UCF1”
How deep learning extracts and learns leaf features for plant classification (2017, background): “Here, we first Dataset Quantity of images Number of categories MS COCO [51] 328k (2.5 million labeled instances) 91 Places2 [52] 8.3 million 365 Sport-1M [53] 1 million 487 Visual Genome QA [54] 1.7 million questions/answer pairs – ILSVRC 2010 [55] 1.4 million 10 0 0 PlantClef2015 dataset [6] 113,20”
Deep Temporal Linear Encoding Networks (2016, methodology): “%) performance comparison of the VGG-16 spatial ConvNets using 3 segments, when combined with context information pre-trained on Places365 [42].”
Fine-tuning Convolutional Neural Networks for Biomedical Image Analysis: Actively and Incrementally (2017, background): “Convolutional neural networks (CNNs) [14] have brought about a revolution in computer vision thanks to large annotated datasets, such as ImageNet [6] and Places [27].”
Globally and locally consistent image completion (2017, methodology): “For the comparison, we have retrained the model of [Pathak et al. 2016] on the Places2 dataset for arbitrary masks for the same number of epochs as our model, and use the best performing model obtained during training.”
Fast Feature Fool: A data independent approach to universal adversarial perturbations (2017, methodology): “Table 4 shows the fooling rates on ILSVRC trained networks, where we used Places-205 data to generate the perturbation for UAP [13] and Table 5 shows the reverse scenario.”