---
title: 'BI-Scan: Efficient Bidirectional Interaction'
url: https://www.emergentmind.com/topics/bidirectional-interaction-scan-bi-scan
type: topic
---

# BI-Scan: Efficient Bidirectional Interaction

Bidirectional Interaction Scan (BI-Scan) is a scanning mechanism introduced in Bidirectional Interaction Mamba (BIM) for multi-task dense prediction. Its purpose is to provide sufficient cross-task interaction without incurring the quadratic cost of fully pairwise exchange. BI-Scan does this by converting task-refined decoder features into bidirectional sequences, processing them with state-space modeling in both task-first and position-first orders, and reintegrating the resulting shared interaction signal through gated fusion. Within the BIM formulation, it is the principal mechanism for preserving the integrity of task-specific information during interaction while maintaining linear complexity in sequence length [2508.20376].

## 1. Problem formulation and design objective

The BI-Scan mechanism is motivated by a trade-off identified in multi-task dense prediction. Pair-wise interaction methods preserve task information well because every task can directly access the others, but they scale as \(O(n^2)\) with task number. Bridge-based linear methods reduce complexity to \(O(n)\), but they do so by compressing all task information into a small shared bridge, which loses details and weakens interaction quality. BI-Scan is proposed to resolve this trade-off by keeping the complexity linear while preserving the integrity of task-specific information during interaction [2508.20376].

In the BIM prediction pipeline, the encoder-decoder backbone is expressed as
\[
\{F_1,F_2,\ldots,F_T\}=\Theta(\Phi(\textnormal{I})),\quad \hat{Y}_{t}=H_t(F_t)=\mathcal{P}_t(\Psi(F_t)).
\]
Within this pipeline, BI-Scan operates on task-specific normalized decoder features \(\{\mathcal{F}^{s}_{t,N}\}_{t=1}^{T}\). The mechanism is explicitly framed as an alternative to interaction schemes that either pay quadratic cost for direct exchange or accept information bottlenecks through bridge compression. A plausible implication is that BI-Scan is intended not merely as a more efficient interaction layer, but as a reformulation of cross-task communication into a scan-compatible sequence modeling problem.

## 2. Bidirectional sequence construction

BI-Scan begins from decoder-stage task-refined features, denoted \(\{\mathcal{F}^{s}_{t,N}\}_{t=1}^{T}\), and converts them into scanable sequences. The features are split along the channel dimension into two halves. One half is sent through a Forward Scan. The other half is reversed along the task dimension and then processed by a Backward Scan. This produces two complementary directional views over the same task set, increasing the chance that interactions are captured from different contextual perspectives [2508.20376].

The forward and backward branches are not treated as independent outputs. The output of the forward scan is later concatenated with the reversed backward output to form the refined shared interaction representation \(\mathcal{F}_{sh}^{s}\). This construction is central: BI-Scan does not collapse all tasks into a single compact bridge representation before interaction. Instead, it serializes task features directly, allowing Mamba or SSM dynamics to operate over long global sequences.

The bidirectional design therefore has two distinct meanings. First, it scans the same task collection in two opposing task-order directions. Second, it preserves separate channel subspaces for those two directions before concatenation. This suggests that bidirectionality in BI-Scan is not a cosmetic analogue of bidirectional sequence encoders, but a structural device for retaining complementary cross-task context.

## 3. Task-first and position-first scan modes

BI-Scan contains two complementary scan modes inside each direction: Task-First and Position-First. They differ in how the multi-task feature tensor is serialized before being passed through the linear-time SSM or Mamba block [2508.20376].

In Task-First mode, task ordering is prioritized before spatial ordering. The first step is pattern-guided feature serialization, denoted \(\mathcal{U}_1\), which serializes each task’s spatial map according to a scan pattern, yielding \(T\) subsequences of length \(H\times W\). The second step is task order-guided aggregation, denoted \(\mathcal{O}_1\), which concatenates those \(T\) subsequences along the sequence dimension in a specified task order. The resulting sequence is
\[
S_1=\mathcal{O}_1(\mathcal{U}_1(\mathcal{F}_1^{s},\mathcal{F}_2^{s},\ldots,\mathcal{F}_T^{s})) \in \mathbb{R}^{(T\times (H \times W)) \times \frac{C}{2}}.
\]
This global sequence is then processed by the SSM to produce cross-task interaction information.

In Position-First mode, the emphasis shifts from task-centric ordering to spatial alignment across tasks. The task-first output is first restored to feature maps \(\{\bar{F}_t^s\}_{t=1}^T\in \mathbb{R}^{H\times W\times \frac{C}{2}}\). Then task order-guided feature serialization, \(\mathcal{O}_2\), gathers tokens at the same spatial location across tasks, and pattern-guided subsequence aggregation, \(\mathcal{U}_2\), concatenates them according to the scan pattern. The resulting sequence is
\[
S_2=\mathcal{U}_2(\mathcal{O}_2(\bar{F}_1^{s},\bar{F}_2^{s},\ldots,\bar{F}_T^{s})) \in \mathbb{R}^{((H \times W)\times T) \times \frac{C}{2}}.
\]

These two orderings are unified inside a linear-time SSM or Mamba framework. Outputs from all scan patterns are fused by element-wise summation to produce the forward result \(\mathcal{F}_{for}^{s}\). The backward scanning branch is then reversed and concatenated with the forward output. BI-Scan thus combines forward and backward directions, task-first and position-first orderings, and multiple scan patterns within one coherent architecture. A common simplification is to treat BI-Scan as only a bidirectional scan; more precisely, its distinguishing feature is the joint use of bidirectionality and dual serialization order.

## 4. Reintegration by gated fusion

The shared interaction representation produced by BI-Scan is not mixed back into task features by unconditional addition. The BIM formulation uses a BCFR fusion step:
\[
F_t^s=\mathcal{F}^{s}_{t}+\mathcal{G}^{s}_{t}\times\mathcal{F}^{s}_{sh}[t]+(1-\mathcal{G}^{s}_{t})\times \mathcal{F}^{s}_{t,N}.
\]
Here, \(\mathcal{G}^{s}_{t}\) is a learned gate from normalized task features, and \(\mathcal{F}^{s}_{sh}[t]\) is the part of the shared interaction feature assigned to task \(t\). This gate allows each task to selectively absorb shared information rather than uniformly mixing all shared content back into every task stream [2508.20376].

This gating is important for interpreting BI-Scan. The mechanism is not only about constructing richer shared interaction states; it also regulates task-specific uptake of that information. In that sense, BI-Scan plus gated reintegration functions as a controlled information exchange scheme rather than a simple shared-feature broadcast.

The broader BIM architecture places BI-Scan alongside Multi-Scale Scan (MS-Scan). Their roles are differentiated. BI-Scan focuses on cross-task interaction. MS-Scan focuses on multi-scale scene modeling within a task. MS-Scan tokenizes features at multiple spatial scales, processes each scale with SS2D scanning, and fuses the results. The paper states that MS-Scan can also be inserted into the BI-Scan framework without increasing computational complexity. This suggests a separation of concerns: BI-Scan governs how tasks interact, whereas MS-Scan governs what spatial granularity is emphasized during representation learning.

## 5. Empirical characteristics on NYUD-V2 and PASCAL-Context

The reported experiments on NYUD-V2 and PASCAL-Context are used to support the claim that BI-Scan improves cross-task interaction quality while preserving efficiency [2508.20376].

On NYUD-V2, the ablation reported for the baseline MTMamba gives Semseg \(55.82\), Depth \(0.5066\), Normal \(18.63\), Boundary \(78.70\), and MTL gain \(+2.38\). Adding BI-Scan alone yields Semseg \(57.11\), Depth \(0.4856\), Normal \(18.66\), Boundary \(78.90\), and MTL gain \(+4.02\). The full BIM model yields Semseg \(57.40\), Depth \(0.4733\), Normal \(18.55\), Boundary \(78.72\), and MTL gain \(+4.83\). These numbers support the specific claim that BI-Scan alone already improves over the baseline, while the full BIM configuration further improves depth.

The ablation analysis also reports several architectural effects. Bidirectional scanning beats unidirectional scanning. Task-First and Position-First each help. Using both together works best. Task order is fairly robust, with random ordering hurting slightly but not catastrophically. The appendix further notes that TF \(\rightarrow\) PF is better than PF \(\rightarrow\) TF, indicating that the order of the two scan modes matters.

On PASCAL-Context, BIM outperforms MTMamba across all tasks: semantic segmentation \(81.25\) versus \(81.11\), parsing \(73.15\) versus \(72.62\), saliency \(84.14\) versus \(84.14\), normal \(14.13\) versus \(14.14\), and boundary \(79.05\) versus \(78.80\). The strongest improvements are reported in parsing and boundary prediction.

The qualitative results are described as showing clearer boundary and detail recovery and better task-specific focus. Visual attention analysis is reported to show that BI-Scan suppresses irrelevant spatial responses, amplifies task-critical regions, and improves structural separation in difficult scenes. A particularly noted benefit is improved separation of architecturally distinct elements, indicating that bidirectional interaction helps the model avoid conflating nearby structures.

## 6. Complexity, efficiency, and scope of the term

BI-Scan is described as providing a richer interaction path than bridge compression while maintaining linear complexity in sequence length. The empirical efficiency argument in the appendix is that model cost grows linearly with task count. In the reported experiment, moving from \(2\) tasks to \(3\) tasks and then \(4\) tasks increased FLOPs by a roughly constant increment, exemplified by \(+86\) GFLOPs per added task [2508.20376].

This efficiency claim should be interpreted narrowly. BI-Scan does not eliminate the cost of global interaction; it restructures that interaction into scanable sequences that can be modeled by SSM or Mamba dynamics with linear scaling. Its advantage therefore lies in preserving direct access to task features without resorting to a small shared bridge, rather than in approximating full pairwise interaction with explicit attention matrices.

The term itself also requires domain-specific precision. In the multi-task dense prediction literature, BI-Scan refers to the Bidirectional Interaction Scan mechanism in BIM [2508.20376]. Closely related phrases appear in other fields, but they denote different methods. “Bidirectional quantitative scattering microscopy” refers to BiQSM, a microscopy system combining forward scattering and backward scattering rather than a multi-task interaction module [2503.14818]. “Aligned Hierarchical Bidirectional Scan” refers to AHBS in ABMamba, a video temporal modeling module for captioning rather than dense prediction [2604.08050]. A bidirectional holographic photothermal microscope has also been used to compare forward-scattering and backward-scattering geometries and to introduce internal forward scattering, but that work does not define BI-Scan as a dense-prediction interaction mechanism [2607.03170]. This suggests that BI-Scan is best treated as a specific mechanism within BIM, not as a generic label for all bidirectional scan-based systems.

## 7. Conceptual significance within multi-task dense prediction

Within BIM, BI-Scan is the mechanism that operationalizes “sufficient cross-task interaction” under a linear-complexity constraint [2508.20376]. It does so through four coupled design choices: direct serialization of task features rather than bridge compression, complementary forward and backward task-direction scans, complementary Task-First and Position-First sequence orderings, and gated task-wise reintegration.

This combination addresses two failure modes identified by the paper. The first is the computational burden of pair-wise interaction. The second is the information degradation caused by compressing heterogeneous task information into a small shared bridge. BI-Scan is intended to avoid both by preserving task-specific feature integrity throughout the interaction process and by making the scan order itself part of the modeling strategy.

A plausible implication is that BI-Scan represents a broader methodological pattern in SSM-based dense prediction: cross-task interaction can be reformulated as structured sequence construction, with the quality of interaction determined not only by the backbone dynamics but also by how tasks and spatial positions are serialized before scanning. In the specific formulation of BIM, that principle is realized through bidirectional task traversal, dual scan orderings, and selective fusion, yielding a cross-task interaction mechanism that is both scalable and information-preserving.

Source: https://www.emergentmind.com/topics/bidirectional-interaction-scan-bi-scan