Learning person-object interactions for action recognition in still images

By Vincent Delaitre, Josef Sivic, Ivan Laptev. 2011

158 citations recorded by the source corpus.

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

Joint Discovery of Object States and Manipulation Actions (2017, background): “Modeling co-occurrences of objects and actions have shown benefits for recognizing actions in [11, 19, 26, 35, 47].”
Weakly-Supervised Learning of Visual Relations (2017, background): “This task was studied for the special case of human-object interactions [8, 9, 17, 38, 39, 48, 49, 50], where the triplet is in the form ( person, action, object ) .”
Anticipating Human Activities Using Object Affordances for Reactive Robotic Response (2013, background): “Another approach is to represent activities as collections of semantic attributes [18], [19], [20], [21].”
Learning to Track for Spatio-Temporal Action Localization (2015, neutral): “Action recognition and localization in still images [3] is an extreme example along the first line, where local detectors are trained e.”
HICO: A Benchmark for Recognizing Human-Object Interactions in Images (2015, background): “detecting human body parts) by deriving feature representations based on human poses [11, 38, 21, 30, 32, 31], whereas others focus more on spatial relations between humans and objects [34, 33, 23, 9, 5, 4].”
On Recognizing Actions in Still Images via Multiple Features (2012, background): “Object-person interactions are explored in other works such as [2, 3, 13, 14].”
Learning Human-Object Interactions by Graph Parsing Neural Networks (2018, methodology): “Early work in HOI understanding studied Bayesian model [15, 16], utilized contextual relationship between human and objects [47–49], learned structured representations with spatial interaction and context [8], exploited compositional models [9], or referred to a set of HOI exemplars [19].”
A survey on still image based human action recognition (2014, methodology): “In their later methods [30,38], the model was extended to deal with any number of objects, instead of being limited to the interactions between one person and one object.”
Attentional Pooling for Action Recognition (2017, background): “Pose: There have also been previous works in incorporating human pose into action recognition [8, 10, 60].”
Joint action recognition and pose estimation from video (2015, background): “For certain actions defined by specific geometric configuration of body parts, pose estimation from a single image may be sufficient for action recognition [15, 28, 5, 32].”
First Person Action-Object Detection with EgoNet (2016, background): “n pose and objects for activity recognition. Some approaches also used low level bag-offeature models to learn the spatial relationship between objects and activities from a single third-person image [10]. Conversely, the activity can provide a functional cue to recognize objects [48, 14, 24]. Such ”
“Important stuff, everywhere!” Activity recognition with salient proto-objects as context (2014, background): “Approaches that solely use image features to automatically retrieve object information mostly rely on trained detectors [16, 11, 35, 25, 27, 6].”
Learning to Detect Human-Object Interactions (2017, methodology): “Results produced in these works were evaluated on either action classification [34, 33, 6, 25, 4, 5, 28, 13], object detection [34], or human pose estimation [34, 5]; none of them were directly evaluated on HOI detection.”
Watching Unlabeled Video Helps Learn New Human Actions from Very Few Labeled Snapshots (2013, methodology): “In particular, “poselets” [4]—local part-based features mined for their consistency with fragments of body pose—have proven to be a promising representation [33, 17, 7], as well as high-level descriptors that also incorporate interactions with objects [8, 34, 7, 36].”
Representing Videos Using Mid-level Discriminative Patches (2013, background): “Furthermore, these patches can be used as a discriminative vocabulary for action classification where they demonstrate stateof-the-art performance on UCF50 and Olympics datasets.”
Action Recognition with Exemplar Based 2.5D Graph Matching (2012, background): “The idea of using human poses for action recognition has been studied in some previous work which either detect local pose features [4, 6] or model the spatial configuration between human body parts and objects [2, 3, 7, 8].”
A similarity measure for analyzing human activities using human-object interaction context (2014, methodology): “OVERVIEW OF EXISTING METHODS Recently, human-object interaction cues have received a lot of attention when designing systems for understanding human actions in a still image [5, 8-10].”