3D intent grounding stops matching verbs to objects: CoCR splits vague needs into a functional chain, reaching 45.91 Top-1@0.5 on the Intent3D test set
Synopsis
In 3D intention grounding, a user says something abstract like "I need back support" without naming an object category, and prior methods matched verbs to objects directly, leaving it unclear whether a candidate truly meets the implied function. CoCR first decomposes the intent into ordered functional requirements, then builds an intent–function–object dependency graph and aligns function-aware representations with point-cloud geometric-semantic features through cross-attention for bidirectional verification. On the Intent3D test set, the full model reaches 45.91 Top-1@0.5 and 30.28 AP@0.5, above IntentNet's 42.28 and 27.60.
Interpretation
3D intention grounding maps abstract intents that name no object category, such as "I need back support," onto concrete objects in a 3D scene, and CoCR routes this through an explicit intermediate layer of functional requirements instead of comparing intent text with candidate objects directly. The earlier IntentNet matched targets through verb–object semantic correspondence; the model could align words like "support" and "back" but could not say why a candidate satisfies the intent, so sofas, pillows and chairs were easily confused in open scenes. On the Intent3D test set the full CoCR reaches 45.91 Top-1@0.5 and 30.28 AP@0.5, against 42.28 and 27.60 for IntentNet; on the validation set Top-1@0.5 rises from 40.83 to 43.24.
The method splits one intent into several functional requirements, links those requirements as intermediate nodes to candidate objects, and then uses point-cloud features to correct the reasoning chain in return. The decomposition is done by a T5-small fine-tuned on the training split, which follows a template of identifying core demands, inferring required functional properties and summarizing functional requirements; the generated steps are projected onto a functional vocabulary built from training-split intent annotations and manually verified, rather than used as free text. Dependency weights between functional requirements and candidate objects are initialized from text-embedding similarity, and weak links below a threshold of 0.2 are masked at inference; cross-attention uses candidate object features as queries and function-aware representations as keys and values, a dependency-weighted bidirectional consistency constraint pulls only functionally relevant object–requirement pairs together, and the graph is updated for two iterations.
Perspective
The demonstrated scope is the ScanNet-derived Intent3D dataset (44,990 intention texts across 1,042 scenes) plus 3D visual grounding evaluation on Nr3D, Sr3D and ScanRefer; the model trains for about 40 hours on four 24GB RTX 3090 GPUs and remains deployable on a single GPU, which suits researchers and engineers working on intent understanding and 3D localization in embodied interaction.
Here "causal" denotes task-level functional causal dependencies, not formal causal identification under a structural causal model; the functional vocabulary is built from training-split intent annotations and manually verified, the parser is a fine-tuned T5-small, more than four reasoning steps introduce redundancy and may degrade performance, and more than two graph update iterations bring marginal benefit and may homogenize object representations through repeated fusion.
