EdgeSGG: Edge-Centric Scene Graph Generation

• EdgeSGG is an edge-centric paradigm for scene graph generation that treats relationships as independent entities with rich geometric and semantic features.
• It employs novel architectures like EdgeGCN, DualMPNN, and LineGNN to evolve node and edge features through dedicated message passing streams.
• Empirical evaluations on datasets such as 3RScan, Visual Genome, and OpenImages demonstrate improved recall, precision, and overall performance.

EdgeSGG denotes a family of edge-centric formulations in which relations are treated as first-class computational entities rather than as by-products of node-centric reasoning. In scene graph generation, this usage appears most explicitly in 3D point-based scene graph analysis, where multi-dimensional edge features are explicitly evolved alongside node features (Zhang et al., 2021), and in semantic scene graph generation, where an edge dual scene graph converts original relations into nodes for relation-centric message passing (Kim et al., 2023). A closely related 3D scene graph prediction line performs edge-centric relational reasoning on a line graph and then fuses enriched relation features back into an object-centric graph (Ma et al., 19 Nov 2025). The terminology is not fully standardized: in Gaussian splatting, “EdgeSGG” is described as likely referring to EGGS, an edge-guided loss for radiance fields rather than a scene graph model (Gong, 2024).

1. Terminology and conceptual core

The central idea of EdgeSGG is a shift from object-centric to relation-centric computation. Conventional scene graph generation typically represents objects as primary nodes and predicts predicates from node pairs, so edges are often secondary, implicit, or reduced to shallow classifiers over concatenated node embeddings. EdgeSGG reverses that priority: edges become explicit, high-dimensional, and updateable carriers of semantics, geometry, and contextual structure. In this view, a scene graph is not merely a set of objects with labeled links, but a coupled system in which node and edge representations co-evolve.

Across the literature, this edge orientation is realized in three closely related ways. One line maintains the original scene graph but equips it with explicit edge tensors and an edge evolution branch, as in the 3D point-based $\mathbf{SGG_{point}}$ framework with EdgeGCN. A second line constructs a dual graph in which original relations become nodes, enabling message passing over relation neighborhoods. A third line forms a line graph from the primitive graph, so that inter-relation dependencies can be modeled directly before being injected back into an object-centric graph. In all three cases, the architectural consequence is the same: relationship reasoning is no longer post-hoc classification, but a dedicated representational process (Zhang et al., 2021).

A plausible implication is that EdgeSGG is better treated as a design paradigm than as a single standardized architecture. The shared design commitments are explicit edge features, edge-conditioned node updates, node-conditioned edge updates, and graph transformations—such as dual graphs or line graphs—that make relation-to-relation interactions computationally accessible.

2. EdgeSGG in 3D point-based scene graph generation

In the 3D point-based formulation, the input scene is a point cloud

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$

and the goal is to produce a semantic scene graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$, where nodes correspond to object instances and directed edges correspond to inter-object relationships. The framework $\mathbf{SGG_{point}}$ is organized into three sequential stages: $\mathrm{Construction}_{\mathbf{SG}}$, $\mathrm{Reasoning}_{\mathbf{SG}}$, and $\mathrm{Inference}_{\mathbf{SG}}$. A shared backbone $\mathcal{F}_{\mathcal{B}}(\cdot)$, instantiated as PointNet or DGCNN, extracts point-wise features $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$, with $\mathcal{C}_{input}=9$ and $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$0. Instance-wise symmetric pooling then yields node features $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$1, while edge features are explicitly parameterized as

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$2

with directed edge feature

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$3

The reasoning stage is implemented by EdgeGCN, which replaces a purely node-centric GCN with two coupled streams: node evolution and edge evolution. Node evolution is modulated by an edge-driven attention mask $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$4, derived from aggregated incoming and outgoing edge features, while edge evolution is modulated by a node-driven attention mask $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$5, derived from the evolved source and target node states. The node update takes the form

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$6

while the edge stream is a two-layer MLP gated by $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$7: $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$8 This “twinning” coupling preserves independent evolution paths for nodes and edges while allowing each stream to modulate the other.

Inference uses residual fusion,

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$9

followed by NodeMLP and EdgeMLP with a joint loss

$\mathcal{G}=(\mathcal{V},\mathcal{E})$0

On 3RScan + 3DSSG, the reported gains include Object R@10 of $\mathcal{G}=(\mathcal{V},\mathcal{E})$1 versus $\mathcal{G}=(\mathcal{V},\mathcal{E})$2 for the baseline, Predicate F1@5 of $\mathcal{G}=(\mathcal{V},\mathcal{E})$3 versus $\mathcal{G}=(\mathcal{V},\mathcal{E})$4, and Triplet R@100 of $\mathcal{G}=(\mathcal{V},\mathcal{E})$5 versus $\mathcal{G}=(\mathcal{V},\mathcal{E})$6. The ablations further show that Vanilla EdgeGCN improves both node and edge recognition relative to shared-feature baselines, that $\mathcal{G}=(\mathcal{V},\mathcal{E})$7 primarily benefits node recognition, that $\mathcal{G}=(\mathcal{V},\mathcal{E})$8 strongly benefits edge recognition, and that using both attentions yields the best joint performance. On SUNCG, $\mathcal{G}=(\mathcal{V},\mathcal{E})$9 matches or surpasses SceneGraphNet in living, office, and bath categories while remaining on par in bed scenes (Zhang et al., 2021).

3. Edge dual scene graphs and DualMPNN

A second major EdgeSGG formulation is built around an edge dual scene graph. Starting from an image $\mathbf{SGG_{point}}$0, an object detector

$\mathbf{SGG_{point}}$1

produces object features, bounding boxes, and class logits or labels. A primitive scene graph

$\mathbf{SGG_{point}}$2

is then formed, where $\mathbf{SGG_{point}}$3 and $\mathbf{SGG_{point}}$4 are subject and object node sets and $\mathbf{SGG_{point}}$5 is the directed relation feature between them. The edge dual scene graph $\mathbf{SGG_{point}}$6 converts original relations into dual nodes: $\mathbf{SGG_{point}}$7 and introduces dual edges whenever two original relations share an object: $\mathbf{SGG_{point}}$8 For a fully connected primitive graph with $\mathbf{SGG_{point}}$9 nodes and $\mathrm{Construction}_{\mathbf{SG}}$0 edges, the dual graph has $\mathrm{Construction}_{\mathbf{SG}}$1 nodes and

$\mathrm{Construction}_{\mathbf{SG}}$2

dual edges.

DualMPNN learns on both $\mathrm{Construction}_{\mathbf{SG}}$3 and $\mathrm{Construction}_{\mathbf{SG}}$4. Its object-centric branch updates relation features according to

$\mathrm{Construction}_{\mathbf{SG}}$5

where

$\mathrm{Construction}_{\mathbf{SG}}$6

Its relation-centric branch initializes from shared-object features,

$\mathrm{Construction}_{\mathbf{SG}}$7

and propagates through the dual graph by

$\mathrm{Construction}_{\mathbf{SG}}$8

The two streams are aggregated by concatenation,

$\mathrm{Construction}_{\mathbf{SG}}$9

and optimized with

$\mathrm{Reasoning}_{\mathbf{SG}}$0

This formulation is evaluated on Visual Genome and OpenImages V6 under PredCls, SGCls, and SGGen. On Visual Genome, the reported EdgeSGG results include PredCls mR@50/100 of $\mathrm{Reasoning}_{\mathbf{SG}}$1, SGCls mR@50/100 of $\mathrm{Reasoning}_{\mathbf{SG}}$2, and SGGen mR@50/100 of $\mathrm{Reasoning}_{\mathbf{SG}}$3; on OpenImages, the method reports mR@50 of $\mathrm{Reasoning}_{\mathbf{SG}}$4, R@50 of $\mathrm{Reasoning}_{\mathbf{SG}}$5, $\mathrm{Reasoning}_{\mathbf{SG}}$6 of $\mathrm{Reasoning}_{\mathbf{SG}}$7, $\mathrm{Reasoning}_{\mathbf{SG}}$8 of $\mathrm{Reasoning}_{\mathbf{SG}}$9, and $\mathrm{Inference}_{\mathbf{SG}}$0 of $\mathrm{Inference}_{\mathbf{SG}}$1. The ablations show that relation-centric learning alone outperforms object-centric learning alone on SGGen mR, and that combining both branches produces the strongest result. The long-tail analysis further reports improved head, body, and tail predicate behavior, while qualitative examples include more specific predicates such as predicting “carrying” instead of a more generic alternative (Kim et al., 2023).

4. Line-graph EdgeSGG and edge-to-object reasoning

A third formulation is the LEO framework, which instantiates edge-centric scene graph prediction through link-guided edge-centric relational reasoning with object-aware fusion. The primitive graph $\mathrm{Inference}_{\mathbf{SG}}$2 is initially fully connected over object instances extracted from a segmented point cloud $\mathrm{Inference}_{\mathbf{SG}}$3 and instance masks $\mathrm{Inference}_{\mathbf{SG}}$4. Object features $\mathrm{Inference}_{\mathbf{SG}}$5 are obtained from multi-scale PointNet; edge features $\mathrm{Inference}_{\mathbf{SG}}$6 are computed from subject-object feature differences and an MLP; and in PredCls the object features may be replaced by label embeddings.

LEO first performs link prediction to suppress irrelevant edges. For each object $\mathrm{Inference}_{\mathbf{SG}}$7, bounding-box and geometric descriptors are embedded as

$\mathrm{Inference}_{\mathbf{SG}}$8

For each pair $\mathrm{Inference}_{\mathbf{SG}}$9, a link feature is constructed as

$\mathcal{F}_{\mathcal{B}}(\cdot)$0

and classified by

$\mathcal{F}_{\mathcal{B}}(\cdot)$1

The positive-link probability is used as a soft edge weight, yielding $\mathcal{F}_{\mathcal{B}}(\cdot)$2. The weighted primitive graph is then transformed into a line graph $\mathcal{F}_{\mathcal{B}}(\cdot)$3, where

$\mathcal{F}_{\mathcal{B}}(\cdot)$4

and

$\mathcal{F}_{\mathcal{B}}(\cdot)$5

Each original relation becomes a node, and two relation-nodes are adjacent if they share an object.

Edge-centric reasoning is then performed by LineGNN. At layer $\mathcal{F}_{\mathcal{B}}(\cdot)$6, relation node $\mathcal{F}_{\mathcal{B}}(\cdot)$7 aggregates messages from neighboring relation nodes: $\mathcal{F}_{\mathcal{B}}(\cdot)$8 with attention

$\mathcal{F}_{\mathcal{B}}(\cdot)$9

and state update

$$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$0

The enriched relation features are then injected back into an object-centric graph, whose node and edge states are refined by GRU updates with messages from incident relations and endpoint objects. Prediction uses object and predicate classifiers, with total objective

$$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$1

On 3DSSG, LEO is reported as model-agnostic and improves both KISGP and 3DHetSGP. For KISGP in PredCls, the reported gains include ngcR@20 from $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$2 to $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$3, ngcR@50 from $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$4 to $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$5, mR@50 from $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$6 to $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$7, and mR@100 from $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$8 to $$\mathcal{X}_{\mathcal{P}}\in\mathbb{R}^{\mathcal{N}\times \mathcal{C}_{point}}$$9. For KISGP in SGCls, ngcR@50 improves from $\mathcal{C}_{input}=9$0 to $\mathcal{C}_{input}=9$1 and mR@50 from $\mathcal{C}_{input}=9$2 to $\mathcal{C}_{input}=9$3. For 3DHetSGP in PredCls, R@20 improves from $\mathcal{C}_{input}=9$4 to $\mathcal{C}_{input}=9$5, ngcR@20 from $\mathcal{C}_{input}=9$6 to $\mathcal{C}_{input}=9$7, and mR@20 from $\mathcal{C}_{input}=9$8 to $\mathcal{C}_{input}=9$9; in SGCls, mR@20 improves from $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$00 to $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$01 and mR@50 from $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$02 to $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$03. The ablations also report best performance at approximately five LineGNN layers and show that inserting LineGNN before the object-centric GNN outperforms post, mix, and parallel integration (Ma et al., 19 Nov 2025).

5. Broader edge-centric graph methodology

The EdgeSGG viewpoint in scene graphs sits within a broader edge-first methodology in graph learning. In edge-featured graph neural architecture search, EGNAS introduces a cell-based search space with two coupled DAGs: an entity-updating graph for node states $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$04 and an edge-updating graph for relation states $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$05. Node updates are edge-guided, while edge updates are node-guided through operations such as Concat, GRU, FiLM, skip-connect, sum, mean, and max. Mixed operations are relaxed by a DARTS-style softmax over candidate operators, and the final discrete architecture is obtained by selecting the highest-weight operations. On ZINC, EGNAS reports MAE $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$06, outperforming GatedGCN(E) at $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$07; on TSP it reports F1 $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$08, exceeding GatedGCN(E) at $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$09; and on CLUSTER it reports $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$10 accuracy (Cai et al., 2021).

A second neighboring line is efficient dynamic graph learning at scale. EDGE models a time-ordered event sequence

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$11

and updates time-dependent node embeddings via GRU-like operators. Its distinctive mechanism is the introduction of selected d-nodes with decoupled embeddings $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$12, enforced through a quadratic penalty

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$13

This shortens the computational DAG and improves parallelism. The method is reported to scale to dynamic graphs with millions of nodes and hundreds of millions of temporal events, and the accompanying discussion explicitly presents it as a foundation for scalable edge-centric systems such as an “EdgeSGG” framework (Chen et al., 2021).

A third adjacent line is the diffusion-based model EDGE for large-graph generation. There, the forward process is discrete edge removal toward the empty graph $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$14, and the reverse process predicts only edges between active nodes. The degree-guided version conditions generation on $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$15 and uses analytically derived node activity probabilities

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$16

Its per-step complexity is

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$17

with total complexity $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$18, where $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$19 is the number of edges and $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$20 the maximum number of active nodes. In the paper’s terminology, this model is described as embodying “edge-based scalable graph generation,” which aligns with an EdgeSGG interpretation outside the scene graph domain (Chen et al., 2023).

6. Nomenclature overlap, misconceptions, and limitations

A notable source of ambiguity is the use of “EdgeSGG” outside scene graph generation. In radiance-field reconstruction, EGGS—“Edge Guided Gaussian Splatting”—introduces an edge-weighted reconstruction loss

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$21

and

$$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$22

The method leaves the Gaussian representation, projection, alpha blending, and standard particle operations unchanged, modifying only the $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$23 term. It is reported to improve PSNR by $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$24 dB on Banana, $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$25 dB on Train, and $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$26 dB on Truck, and the paper explicitly states that “EdgeSGG” likely refers to the same concept in practical usage (Gong, 2024). This usage is structurally unrelated to scene graph generation, but it preserves the broader theme of edge-aware optimization.

A recurrent misconception is that edge-centric models discard node reasoning. The surveyed architectures do not do so. EdgeGCN uses two coupled streams rather than a pure edge-only system; DualMPNN combines object-centric and relation-centric message passing; and LEO performs relation-level reasoning first, then reinjects the resulting edge features into an object-centric graph. EdgeSGG therefore denotes a redistribution of representational authority toward edges, not the elimination of nodes (Zhang et al., 2021).

Another misconception is that edge-centricity is equivalent to adding edge features to an otherwise unchanged GNN. The literature is more specific. EdgeGCN learns multi-dimensional attentional couplings between node and edge streams; DualMPNN requires an explicit edge dual scene graph; and LEO requires link prediction, line graph construction, LineGNN, and object-aware fusion. In each case, the architectural topology—not merely the feature vocabulary—changes materially (Kim et al., 2023).

The principal limitations are likewise structural. In the edge dual scene graph formulation, generating an additional dual graph and running an extra message passing process increases complexity, and the dual graph cardinality can grow as $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$27 in fully connected settings (Kim et al., 2023). In LEO, performance depends on link prediction quality, and dense line graphs increase memory and runtime; the reported overhead includes KISGP inference time from $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$28 to $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$29 ms and memory from $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$30 to $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$31 GB, and for 3DHetSGP from $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$32 to $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$33 ms and $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$34 to $$\mathcal{P} = \{\mathcal{P}_k\}_{k=1}^{\mathcal{N}}, \quad \mathcal{P}\in \mathbb{R}^{\mathcal{N}\times \mathcal{C}_{input}},$$35 GB (Ma et al., 19 Nov 2025). In EGGS, gradient-based edge maps are sensitive to noisy textures, benefits are limited in textureless regions, and cluttered scenes reduce the gain relative to simpler scenes (Gong, 2024).

Taken together, these constraints indicate that EdgeSGG is most effective when relational structure is both semantically important and computationally exploitable: settings with informative edge neighborhoods, shared-object relation clusters, or edge-specific high-frequency cues. This suggests that future standardization of the term, if it occurs, will likely remain anchored in the broader principle common to all of its current uses: edges are not ancillary annotations, but central objects of modeling.