PaperlayerView exact revision on arXiv

Source-linked AI summary

Scene Graph Generation from Objects, Phrases and Region Captions

Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, Xiaogang Wang

arXiv:1707.09700v2cs.CV

TL;DR

Object detection, scene graph generation, and region captioning describe connected scene content at different semantic levels, but weak alignment makes them difficult to learn jointly. MSDN dynamically aligns object, phrase, and caption regions and passes messages across them in one end-to-end network. Joint inference improves all three tasks, with a 3.63%–4.31% scene graph generation advantage over state-of-the-art methods on Visual Genome.

  • Problem

    Weak alignment among object detection, scene graph generation, and region captioning complicates joint learning despite their connections across semantic levels.

  • Method

    MSDN dynamically builds spatial- and semantic-based links among object, phrase, and caption regions, then jointly refines their features through graph-guided message passing.

  • Results

    3.63%–4.31%: MSDN outperforms state-of-the-art methods on scene graph generation on Visual Genome and improves all three tasks through joint inference.

  • Takeaways & Limitations

    Jointly modeling the three semantic levels provides mutual improvements for image understanding within the evaluated tasks.

  • Takeaways & Limitations

    Region captioning can be misled when object and relationship misclassifications cause the caption branch to recognize a region incorrectly.

Abstract

from arXiv · show

Object detection, scene graph generation and region captioning, which are three scene understanding tasks at different semantic levels, are tied together: scene graphs are generated on top of objects detected in an image with their pairwise relationship predicted, while region captioning gives a language description of the objects, their attributes, relations, and other context information. In this work, to leverage the mutual connections across semantic levels, we propose a novel neural network model, termed as Multi-level Scene Description Network (denoted as MSDN), to solve the three vision tasks jointly in an end-to-end manner. Objects, phrases, and caption regions are first aligned with a dynamic graph based on their spatial and semantic connections. Then a feature refining structure is used to pass messages across the three levels of semantic tasks through the graph. We benchmark the learned model on three tasks, and show the joint learning across three tasks with our proposed method can bring mutual improvements over previous models. Particularly, on the scene graph generation task, our proposed method outperforms the state-of-art method with more than 3% margin.

1. Introduction

The paper jointly addresses object detection, scene graph generation, and region captioning across different semantic levels, whose weak alignment complicates unified learning. MSDN aligns these levels with a dynamic graph and feature message passing, yielding mutual improvements, including a 3.63%–4.31% scene-graph-generation gain over state-of-the-art methods on Visual Genome.

  • Motivation: Object detection identifies individual objects, scene graph generation adds their directed relationships, and region captioning describes objects, attributes, and interactions in free-form language.These tasks operate at different semantic levels but describe connected aspects of the same visual scene.
  • Motivation: Weak alignment across the three tasks makes joint learning difficult despite their connections.The paper therefore investigates understanding images through all three levels with a single neural network.
  • Approach: Spatial and semantic correlations let phrases and region captions constrain object existence, attributes, and relationships across semantic levels.For example, phrase and caption information can jointly support recognition of people, a toothbrush, attributes, and interactions.
  • Approach: MSDN jointly refines features across the three tasks using an alignment framework and message passing to exploit complementary effects.The model is an end-to-end network that simultaneously detects objects, recognizes relationships, and predicts captions at salient image regions.
  • Results: 3.63%–4.31%: MSDN outperforms state-of-the-art methods on scene graph generation on Visual Genome.The paper also reports mutual improvements on object detection and region captioning.

2. Related Work

Prior work separately advances object detection, visual relationship detection, region captioning, and multi-task learning. MSDN distinguishes itself by explicitly connecting objects, phrases, and captions through spatial and semantic relationships, integrating scene graphs with region captioning.

  • Object Detection: Faster R-CNN remains a popular object-detection basis because of its strong performance, and MSDN adopts its region-based pipeline.The paper positions object detection as foundational for constructing scene graphs.
  • Visual Relationship Detection: Existing visual relationship methods model object relationships, but prior connections between phrases and captions are not built explicitly.MSDN treats objects, phrases, and region captions as semantic levels and connects them using spatial and semantic relationships.
  • Image Captioning: Existing image and region captioning methods do not explicitly leverage scene graphs, whereas MSDN integrates scene-graph structure into region-captioning features.The paper also states that captioning can provide additional information for scene graph generation.
  • Multi-task Learning: Multi-task learning provides precedent for sharing representations among correlated tasks, while MSDN targets complementary effects across three semantic levels.Its focus is simultaneous modeling of object detection, scene graph generation, and region captioning.

3. Multi-level Scene Description Network

MSDN is an end-to-end network with parallel branches for object detection, phrase detection, and region captioning. It extracts ROI features, dynamically links regions across branches, and refines their features through graph-guided message passing before task-specific predictions.

  • Architecture: MSDN uses shared convolutional layers and three parallel branches for object detection, phrase detection, and region captioning.The architecture is based on a VGG-16 region-based detection pipeline.
  • Processing Pipeline: The pipeline generates proposals, specializes ROI features, constructs a dynamic cross-level graph, and refines features before task-specific outputs.The graph models connections among object, phrase, and caption features using corresponding ROI relationships.
  • Region Features: Phrase regions are formed by grouping object regions into pairs, while object and caption regions are produced by region proposal networks.ROI pooling supplies features for the object, phrase, and caption branches.
  • Feature Refining: After message passing, refined object, phrase, and caption features support object detection, phrase detection, and region captioning, respectively.Message passing is guided by a dynamic graph built from object and caption region proposals.

3.1. Region Proposal

MSDN creates three proposal sets for objects, phrases, and caption regions. Object and caption proposals come from RPNs, while phrase proposals connect ordered pairs of object proposals; shared convolutional layers and suppression reduce proposal complexity.

  • Proposal Sets: Three proposal sets are generated for object regions, phrase regions, and caption regions.These proposals provide the regions used by the three semantic branches.
  • Object Proposals: Object region proposals are generated directly by an RPN.
  • Phrase Proposals: Phrase proposals group N object proposals into N(N −1) ordered object pairs, excluding pairs of identical proposals.The construction fully connects object proposals with directed edges.
  • Caption Proposals: Caption region proposals are generated directly by a separate RPN trained with ground-truth region bounding boxes.
  • Proposal Refinement: The object and caption RPNs share VGG-16 base convolutional layers, while non-maximum suppression separately reduces object and caption ROI sets.Their anchors are obtained by k-means clustering of logarithmic ground-truth box widths and heights.

3.2. Feature Specialization

The model specializes object, phrase, and caption-region features into separate task-specific branches after ROI pooling, while proposals are organized across semantic levels.

  • Feature Specialization: Three ROI sets are processed through ROI pooling and separate fully connected layer sets for object, phrase, and caption branches.Each branch uses two 1024-dimensional fully connected layers in the implementation.
  • Feature Specialization: Figure 3 organizes object proposals at the bottom, phrase proposals in the middle, and caption-region proposals at the top.The figure also depicts graph connections between these proposal types; some phrase boxes are omitted.

3.3. Dynamic Graph Construction

MSDN dynamically builds a graph whose topology varies by image, connecting object, phrase, and caption proposals through semantic and spatial relationships.

  • Dynamic Graph Construction: The connection graph is dynamically constructed because different input images have different connection topologies among their regions of interest.The graph is based on semantic and spatial relationships among the ROIs.
  • Dynamic Graph Construction: Phrase proposals connect paired object proposals as directed subject-predicate-object triplets, while captions connect to phrases when they cover at least 0.7 of a phrase proposal.Direct caption-object edges are omitted because captions can connect indirectly through phrase nodes.
  • Dynamic Graph Construction: Figure 4 distinguishes object, phrase, and caption feature-refining nodes and uses different line types for subject-predicate versus predicate-object connections.Arrows indicate message-passing direction.
  • Dynamic Graph Construction: The graph contains specialized-feature nodes and undirected caption-phrase edges plus directed subject-predicate and object-predicate edge sets.The subject and object labels identify the two roles associated with each phrase.

3.4. Feature Refining

Feature refining passes messages among object, phrase, and caption nodes in parallel, using merge-and-refine operations that can be repeated iteratively.

  • Feature Refining: Message passing is divided into parallel object-refining, phrase-refining, and caption-refining steps, and the procedure can be applied iteratively.The steps update features through graph edges.
  • Object Feature Refining: Object nodes separate incoming phrase features by subject-predicate and predicate-object connection types before refinement.The two connection types are merged separately and then used to refine the object feature.
  • Object Feature Refining: A gate function weights and averages multiple phrase features because different phrases have different importance for refining an object.The gate determines the weights applied during feature merging.
  • Object Feature Refining: Additional non-shared fully connected layers transform phrase features for subject-predicate and predicate-object message paths.The transformations adapt phrase features before they are incorporated into object refinement.
  • Phrase and Caption Refining: Phrase nodes receive information from their subject and object nodes, while caption nodes receive merged information from connected phrase nodes.Phrase and caption refinement follow the same merge-and-refine paradigm.
  • Feature Refining: The refining structure updates object, phrase, and caption features by absorbing supplementary information from connected nodes.

3.5. Scene Graph Generation

Scene graph generation represents objects and pairwise relationships in a matrix, then forms graph edges from non-background objects with relevant predicates.

  • Scene Graph Generation: Message passing between object and phrase nodes supports direct prediction of object categories and corresponding pairwise relationship categories.
  • Scene Graph Generation: A scene-graph matrix places object predictions on the diagonal and relationship phrases between object pairs at off-diagonal positions.The ith object occupies element (i, i), while element (i, j) represents the relationship from object i to object j.
  • Scene Graph Generation: Objects and predicates are predicted from refined object and phrase features, using background and irrelevant classes to suppress nonexistent nodes or relations.

3.6. Region Caption Generation

Region caption generation uses an LSTM-based language model to turn region features into free-form sentences. The model encodes image and word information, models sequence dependencies, and decodes words until sentence completion.

  • An LSTM-based language model generates natural sentences from region features containing objects, interactions, attributes, and scene information.The region features are passed to the language model for free-form regional descriptions.
  • The captioning model contains image and word encoders, a two-layer LSTM, and a word decoder.The image encoder maps visual features into the word-feature domain, while the word encoder produces word embeddings.
  • At generation time, the encoded image feature starts the sequence, followed by a start token and recursively predicted words until an end token or maximum length.

4. Experiment

Experiments on a cleansed Visual Genome dataset evaluate MSDN across scene graph generation, object detection, and region captioning. Ablations indicate that cross-level message passing and caption supervision improve performance, while excessive refinement can hinder training.

  • Dataset and evaluation: Experiments use Visual Genome relationship and region-description subsets sharing images but targeting different tasks.
  • Dataset and evaluation: The cleansed dataset contains 95,998 images, with 70,998 for training and 25,000 for testing.Preprocessing selects frequent object and predicate categories, normalizes annotations, and removes small boxes and regions.
  • Dataset and evaluation: Rec@50 and Rec@100 measure the fraction of ground-truth instances found among the top-K predictions because relationship annotations are incomplete.Mean average precision would penalize positive but unlabeled relationship predictions.
  • Component analysis: 5.34% ∼6.69%: feature refining message passing improves scene graph generation over a baseline without message passing or caption branching.The comparison attributes the gain to leveraging connections between object and phrase nodes.
  • Component analysis: 0.47% ∼0.84%: adding a caption branch without caption supervision improves performance over the object-phrase message-passing model.The authors state that this gain is more likely due to the additional parameters introduced by the caption branch.
  • Component analysis: 2.03% ∼2.64%: region-caption supervision improves over the unsupervised caption-branch setting through complementary learned features.The supervised caption branch learns specialized features that provide extra information to other branches.
  • Component analysis: Two feature-refining iterations are optimal for scene graph generation; a third iteration decreases performance by 0.21% ∼0.27% because training becomes harder.Two iterations add a 0.33% ∼0.49% gain over one iteration by indirectly connecting captions and objects.
  • Object detection: MSDN outperforms Faster R-CNN and a three-branch baseline for object detection by using phrase and caption context as complementary visual cues.

5. Conclusion

The paper jointly models object detection, visual relationship detection, and region captioning with MSDN, using a dynamically constructed graph to align regions across semantic levels. Experiments show that joint inference improves all three tasks.

  • MSDN jointly models object detection, visual relationship detection, and region captioning in a single end-to-end deep neural network.
  • A dynamically constructed graph establishes links among regions with different semantic meanings and aligns features from the three tasks.
  • Joint inference with MSDN brings improvement in all three tasks.
Loading 1707.09700v2…