End-to-end integration of a Convolutional Network, Deformable Parts Model and non-maximum suppression

By Li Wan, David Eigen, Rob Fergus. 2014

100 citations recorded by the source corpus.

Deformable Parts Models and Convolutional Networks each have achieved notable performance in object detection. Yet these two approaches find their strengths in complementary areas: DPMs are well-versed in object composition, modeling fine-grained spatial relationships between parts; likewise, ConvNets are adept at producing powerful image features, having been discriminatively trained directly on the pixels. In this paper, we propose a new model that combines these two approaches, obtaining the advantages of each. We train this model using a new structured loss function that considers all bounding boxes within an image, rather than isolated object instances. This enables the non-maximal suppression (NMS) operation, previously treated as a separate post-processing stage, to be integrated into the model. This allows for discriminative training of our combined Convnet + DPM + NMS model in end-to-end fashion. We evaluate our system on PASCAL VOC 2007 and 2011 datasets, achieving competitive results on both benchmarks.

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

Object Detection Networks on Convolutional Feature Maps (2015, methodology): “extractor is “upgraded” to a pre-trained deep Con25 vNet, but the classifier is left as a traditional model, 26 such as a DPM [8], [9], [10] or a boosted classifier 27”
Learning Non-maximum Suppression (2017, background): “One idea is to include GreedyNMS at training time [33, 12], making the classifier aware of the NMS procedure at test time.”
CRF-CNN: Modeling Structured Information in Human Pose Estimation (2016, methodology): “The other is to model structures among features and outputs, either as post processing Deng et al. (2014); Chen & Yuille (2014) or as extra information to guide the learning of CNN Zheng et al. (2015); Tompson et al. (2014); Wan et al. (2014).”
Object Detection in 20 Years: A Survey (2019, methodology): “Greedy selection is an old fashioned but the most popular way to perform NMS in object detection.”
Modeling local and global deformations in Deep Learning: Epitomic convolution, Multiple Instance Learning, and sliding window detection (2015, methodology): “Aspect ratio variability and search Sliding window detectors have typically substantially lower performance than region-based systems, even when trained with CNN features [13,33,44]; e.”
RIFD-CNN: Rotation-Invariant and Fisher Discriminative Convolutional Neural Networks for Object Detection (2016, methodology): “Since the groundbreaking success of deep convolutional neural networks (CNN) [1] in image classification task [2] on the ImageNet large scale visual recognition challenge (ILSVRC) [3, 4], CNN-based object detection methods have recently attracted a great deal of research interest and have achieved s”
WILDCAT: Weakly Supervised Learning of Deep ConvNets for Image Classification, Pointwise Localization and Segmentation (2017, unknown): “While [22, 60] are direct generalizations of DPM to CNN, R-FCN [10] improves performances by explicitly learning several part models and using a part-based pooling of features designed for accurate spatial localization and directly inserted at the top of the network.”
End-to-End Training of Object Class Detectors for Mean Average Precision (2016, methodology): “ubic complexity in the number of training samples, making it intractable when there are tens of classes and thousands of windows (e.g. in PASCAL VOC) Unlike most other approaches to object detection, [28] includes NMS at training time as well as test time. They use a deformable parts model over CNN ”
Joint Object and Part Segmentation Using Deep Learned Potentials (2015, background): “In [18], Girshick et. al connected the DPM root filter and part filters with the convolutional filters from CNN, and the distance transform can be regarded as an additional pooling and geometry filtering step.”
Learning Chained Deep Features and Classifiers for Cascade in Object Detection (2017, background): “For example, when the ﬁrst classiﬁer ﬁnds that the object should only be a mammal, then the features and clas-siﬁer at the second stage can focus more on distinguishing speciﬁc class of mammal like horse, antelope and cattle.”
Deep Learning for Generic Object Detection: A Survey (2018, methodology): “In generic object detection, recent work to address geometric transformations and part deformations [49, 84, 184, 199, 270] can be summarized as follows.”
Individualness and Determinantal Point Processes for Pedestrian Detection (2016, background): “The NMS scheme can also be integrated into deep learning models [14].”