Spatial Progressive Augmentation (SPA)

• Spatial Progressive Augmentation (SPA) is a rule-based module in ColLab that refines duplicate referring expressions by applying recursive spatial partitioning.
• It operates after semantic consolidation, using bounding box metadata to assign coarse region labels like center, transition, or edge for disambiguation.
• The module’s progressive refinement effectively isolates same-category instances, though its fixed spatial vocabulary limits broader relational descriptions.

Spatial Progressive Augmentation (SPA) is a rule-based module introduced within ColLab, a collaborative spatial progressive data engine for Referring Expression Comprehension (REC) and Referring Expression Generation (REG). In that setting, SPA is designed to transform semantically plausible but non-unique object descriptions into spatially discriminative referring expressions, especially when an image contains multiple instances of the same category that receive the same base description. Its core operation is to detect duplicated category-description pairs, assign those instances to coarse spatial regions, recursively subdivide ambiguous regions, and rewrite the descriptions with progressively finer spatial terms until each instance is uniquely referable (Zhang et al., 28 Sep 2025).

1. Definition and task setting

In ColLab, REC is defined as the task of identifying the target object or region given an image and a natural-language referring expression, while REG is the reverse task of generating a natural-language expression for a given image region or target object. The paper positions SPA as a response to a specific failure mode in fully automated REC/REG data generation: multimodal models often produce descriptions that are semantically valid but not unique enough for REC, particularly when several objects in one image share the same category and visible attributes (Zhang et al., 28 Sep 2025).

SPA is therefore not a general-purpose language generator. It is a post-generation augmentation module whose function is to increase what the paper calls “spatial expressiveness among duplicate instances.” Concretely, it operates when multiple instances in the same image have the same category and the same generated description. In that circumstance, semantic description alone is insufficient for REC, because the expression does not isolate a single target. SPA addresses that ambiguity by injecting structured spatial cues derived from object geometry (Zhang et al., 28 Sep 2025).

The method is consequently hybrid in a narrow sense: it depends on both language descriptions and object localization metadata. The paper explicitly defines its input as a set

$$S \text{ containing } \{x_{min}, y_{min}, x_{max}, y_{max}, c, d\},$$

where the coordinates specify the bounding box, $c$ is the category, and $d$ is the description. The output is augmented instance descriptions (Zhang et al., 28 Sep 2025).

2. Position within the ColLab pipeline

SPA is the third stage of the ColLab framework. The full pipeline has three stages: scene collection and instance detection, Collaborative Multimodal Model Interaction (CMMI), and Spatial Progressive Augmentation. In the first stage, images are collected from diverse scenes and a detection model extracts candidate object instances; only detections with confidence greater than $0.5$ are kept. In the second stage, each cropped instance is sent to multiple MLLMs, and an LLM merges their outputs into a refined instance description by extracting common characteristics across models. SPA operates only after this semantic consolidation step (Zhang et al., 28 Sep 2025).

This placement is methodologically important. CMMI and SPA solve different problems. CMMI improves semantic quality, consistency, and diversity of per-instance descriptions. SPA intervenes only when those descriptions remain ambiguous after fusion. The paper’s logic is explicit: MLLMs provide rich instance semantics, the LLM distills them into a coherent base expression, and SPA adds spatial specificity only where necessary. This suggests that SPA is best understood as a targeted disambiguation layer, not as a substitute for semantic description generation (Zhang et al., 28 Sep 2025).

Because the trigger condition is duplication of category-description pairs within a single image, SPA is not applied uniformly across all detected objects. It is applied to collision groups: subsets of instances that semantic generation failed to separate. That selective application is central to the design, since it confines spatial rewriting to the cases where uniqueness is otherwise lost (Zhang et al., 28 Sep 2025).

3. Algorithmic mechanism

The procedure begins by grouping instances that share both category and description. The paper formalizes the duplicate-detection step through the condition

$\text{GroupBy}(I_i[c], I_i[d]) > 1.$

If more than one instance has the same category-description pair, the corresponding group is added to the duplicated set $K$ (Zhang et al., 28 Sep 2025).

For each duplicate group $k_i \in K$, SPA initializes a spatial region set $H$ and assigns each instance to a region using its bounding box coordinates $\{x_{min}, y_{min}, x_{max}, y_{max}\}$. The assignment is not relational in the sense of object-to-object reasoning. It is based on a predefined image partition. The paper specifies three region types—center, transition, and edge—and four directional areas—top, bottom, left, and right. Initial spatial terms are therefore structured combinations such as left-transition or bottom-center (Zhang et al., 28 Sep 2025).

After this coarse assignment, SPA checks whether any region still contains multiple duplicate instances. The refinement rule is

$\text{number of instances in } h_j \geq 2.$

When that condition holds, the ambiguous region is subdivided again using the same partitioning scheme. The paper explicitly describes this as a recursive strategy: if multiple instances remain in the same region, that region is partitioned into smaller sections using the same center/transition/edge and directional organization, and the process repeats until each instance is allocated to a distinct subregion (Zhang et al., 28 Sep 2025).

The “progressive” property of SPA is therefore hierarchical spatial refinement rather than a curriculum over epochs or a multi-round prompting loop. Progression means that the method starts with a coarse partition of the image and increases spatial granularity only where ambiguity remains. Once an instance reaches a unique terminal region, its description is updated with the corresponding spatial term. The paper summarizes the final operation as: update instance descriptions with spatial terms based on their final region (Zhang et al., 28 Sep 2025).

4. Spatial vocabulary and representational scope

The spatial cues used by SPA are deliberately limited. The paper supports only three classes of descriptors: region descriptors (center, transition, edge), directional descriptors (top, bottom, left, right), and hierarchical subregions obtained by recursive reapplication of the same partition scheme. It does not describe the method as using object-to-object relations such as “left of the car,” ordinal descriptions such as “second from the left,” depth ordering, or a learned scoring model over relation candidates (Zhang et al., 28 Sep 2025).

That restricted vocabulary has two consequences. First, SPA is computationally simple: it depends only on bounding boxes, duplicate-description filtering, deterministic region assignment, recursive subdivision, and description rewriting. Second, its spatial expressiveness is correspondingly narrow. The paper’s phrase “enhance spatial expressiveness among duplicate instances” refers specifically to region-based positional differentiation within a single image, not to a broad spatial-language planner (Zhang et al., 28 Sep 2025).

The module is therefore best characterized as a region-based recursive spatial partitioning method. Its representational scope is sufficient for the failure mode it targets—same-category instances with identical descriptions in one image—but narrower than general referring-expression systems that model richer relation inventories. A plausible implication is that SPA prioritizes deterministic uniqueness over linguistic variety, although the paper does not formalize that trade-off (Zhang et al., 28 Sep 2025).

5. Illustrative example and empirical status

The paper’s clearest qualitative example involves three traffic lights that initially receive the same description:

“The traffic light in the image is red.”

SPA first partitions the image coarsely. In the example, the green instance is assigned to the lower part of the bottom-center region, while the red and blue instances both fall into the left-transition region. Because the green instance is already unique, it is not refined further. The red and blue instances are then grouped into a new region and recursively repartitioned; after refinement, the red instance is assigned to the left-edge region and the blue instance to the right-edge region (Zhang et al., 28 Sep 2025).

This example is methodologically informative because it shows that refinement is local rather than global. Some instances become uniquely referable at the first partition level, while only the unresolved subset undergoes further subdivision. The paper treats this as evidence that SPA “effectively enhances spatial granularity” and “resolves semantic duplication in multi-instance settings.” It also illustrates the stopping criterion: recursion ends when no occupied region contains more than one instance from the duplicate set (Zhang et al., 28 Sep 2025).

The empirical evidence reported specifically for SPA is primarily qualitative. The paper includes a subsection titled “Advantages Provided by Spatial Progressive Augmentation Module”, but it does not provide a numerical ablation table isolating SPA with metrics such as REC accuracy, REG quality, duplicate-rate reduction, or annotation efficiency. Accordingly, SPA’s contribution is demonstrated through qualitative examples and pipeline-level interpretation rather than a standalone quantitative benchmark (Zhang et al., 28 Sep 2025).

6. Strengths, limitations, and terminological ambiguity

The strengths emphasized by the paper are direct. SPA addresses a concrete failure mode in automatic REC/REG data generation; it requires only bounding boxes and generated descriptions; it is progressive and adaptive in the sense that refinement is applied only when ambiguity remains; and it is compatible with a fully automated pipeline without human supervision (Zhang et al., 28 Sep 2025).

The limitations are also largely implied by the design. SPA is tightly tied to detected image-plane position, so its effectiveness depends on detection quality and on the adequacy of bounding boxes as spatial surrogates. Its spatial vocabulary is limited to region- and direction-based cues. It assumes that duplicate descriptions can be detected at the text level and that image-plane location is sufficient to disambiguate many same-category instances. The absence of a formal quantitative ablation means that the magnitude of SPA’s isolated contribution is not rigorously measured in the paper (Zhang et al., 28 Sep 2025).

The acronym SPA is also notably ambiguous in recent arXiv literature. In “Context-Aware Token Selection and Packing for Enhanced Vision Transformer,” SPA means Select and Pack Attention, a sparse-attention/token-selection mechanism for vision transformers, not Spatial Progressive Augmentation (Zhang et al., 2024). In “SPA: 3D Spatial-Awareness Enables Effective Embodied Representation,” SPA denotes a 3D spatial-awareness framework for embodied representation learning (Zhu et al., 2024). In “Unlocking Patch-Level Features for CLIP-Based Class-Incremental Learning,” SPA refers to Semantic-guided Patch-level Alignment (Sun et al., 13 May 2026). Within this terminological landscape, Spatial Progressive Augmentation refers specifically to the ColLab module for recursive spatial disambiguation of referring expressions (Zhang et al., 28 Sep 2025).

Taken in its own terms, Spatial Progressive Augmentation is a narrowly scoped but well-defined mechanism: after semantic description generation, it detects duplicate category-description pairs, assigns region-based spatial labels, recursively refines ambiguous regions, and rewrites descriptions until each target instance becomes uniquely referable. Its significance lies less in broad spatial modeling than in making fully automated REC/REG data generation operational under duplicate-instance ambiguity (Zhang et al., 28 Sep 2025).