Layered Object Models for Image Segmentation

By Yi Yang, Sam Hallman, Deva Ramanan, Charless C. Fowlkes. 2012

159 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

Panoptic Segmentation (2018, background): “Second, it is interpretable and easy to understand (and does not require solving a complex matching problem as is commonly the case for these types of metrics [13, 49]).”
Recurrent Instance Segmentation (2015, background): “Several works [15,16,17] aim to overcome this problem by inferring and exploiting depth information as a way to order the instances.”
Hypercolumns for object segmentation and fine-grained localization (2014, background): “[42] use figure ground masks associated with DPM detectors [16] to segment out detected objects and reason about depth orderings.”
Pixelwise Instance Segmentation with a Dynamically Instantiated Network (2017, methodology): “This measure was originally proposed in [47], but has not been used since in evaluating instance segmentation systems.”
Simultaneous Detection and Segmentation (2014, methodology): “[32] use object detections from the deformable parts model [13] to segment the image, pasting figure-ground masks and reasoning about their relative depth ordering.”
Attribute2Image: Conditional Image Generation from Visual Attributes (2015, background): “1D representations of images have been studied in the context of moving or still object segmentation [38,23,39,41,11].”
Proposal-Free Network for Instance-Level Object Segmentation (2015, neutral): “Other works have resorted to the object detection task to initialize the instance segmentation and the complex post-processing such as integer quadratic program [34] and probabilistic model [38] to further determine the instance segments.”
Higher Order Conditional Random Fields in Deep Neural Networks (2015, methodology): “Unlike other approaches that reﬁne a bounding-box detection to produce a segmentation [17, 50], this method used detector outputs as a soft constraint and can thus recover from object detection errors.”
Monocular Object Instance Segmentation and Depth Ordering with CNNs (2015, methodology): “Most related to our work are [36, 16, 39] which explicitly reason about class and instance segmentations as well as depth ordering.”
Pixel-Level Encoding and Depth Layering for Instance-Level Semantic Labeling (2016, background): “Using a bank of object detectors as proposals, [32] infers the instance masks via occlusion reasoning based on discrete depth layers.”
Iterative Instance Segmentation (2015, background): “[33] start from object detections from the deformable parts model [10] and paste figure-ground masks for each detected object.”
Instance-Level Segmentation for Autonomous Driving with Deep Densely Connected MRFs (2015, background): “While detection might have been a good proxy when performance was low, recent work has been trying to go beyond simple boxes by providing a detailed segmentation mask for each object instance [12, 30, 28, 13, 29].”
Semantic Instance Segmentation with a Discriminative Loss Function (2017, background): “Two earlier works [43, 36] also use depth information to segment instances.”
Recurrent Pixel Embedding for Instance Grouping (2017, neutral): “Common approaches to instance segmentation ﬁrst generate region proposals or class-agnostic bounding boxes, segment the foreground objects within each proposal and classify the objects in the bounding box [92, 53, 31, 12, 17, 56, 34].”
segDeepM: Exploiting segmentation and context in deep neural networks for object detection (2015, methodology): “Segmentation and detection has also been addressed in a joint formulation in [25] by combining shape information obtained via DPM parts as well as color and boundary cues.”
Weakly- and Semi-Supervised Panoptic Segmentation (2018, background): “However, as noted by several authors [63, 16, 27, 21], the AP r is a ranking metric that does not penalise methods which predict more instances than there actually are in the image as long as they are ranked correctly.”
Beyond Bounding Boxes: Precise Localization of Objects in Images (2015, background): “Yang et al. [61] use figure ground masks associated with DPM detectors [21] to segment out detected objects and reason about depth orderings.”
Efficient inference in occlusion-aware generative models of images (2015, background): “Yang et al. (2012), for example, propose a layered model for segmentation but rely heavily on bounding box and categorical annotations.”
Hough Regions for Joining Instance Localization and Segmentation (2012, background): “[15] propose a layered object model for image segmentation, which defines shape masks and explains the appearance, depth ordering, and labels of all pixels in an image.”