FocusMamba: Efficient RGB-Event Detection

• FocusMamba is an RGB-Event collaborative object detection method that adaptively discards low-information tokens to reduce computation.
• It employs an Event-Guided Multimodal Sparsification module and a Cross-Modality Focus Fusion mechanism to selectively integrate complementary features.
• Empirical results on datasets like DSEC-Det and PKU-DAVIS-SOD demonstrate improved mAP scores with significant FLOPs reduction compared to prior models.

Searching arXiv for the FocusMamba paper and closely related Mamba vision references to ground the article in current literature. to=arxiv_search 大发快三是不是 code 娱乐开号անչிக## search {"query":"Focus Through Motion RGB-Event Collaborative Token Sparsification for Efficient Object Detection FocusMamba arXiv", "max_results": 5} to=arxiv_search 天天彩票怎么 code 】【。】【”】【search {"query":"Mamba architecture vision applications survey arXiv 2025", "max_results": 5} to=arxiv_search വിവരം code 】【。】【search {"query":"(Yang et al., 4 Sep 2025)", "max_results": 5} FocusMamba is an RGB-Event collaborative object detection method introduced in "Focus Through Motion: RGB-Event Collaborative Token Sparsification for Efficient Object Detection" (Yang et al., 4 Sep 2025). It is designed for the setting in which existing RGB-Event detection methods process the low-information regions of both modalities uniformly during feature extraction and fusion, resulting in high computational costs and suboptimal performance. The method addresses this inefficiency through adaptive collaborative sparsification of multimodal features and a Mamba-based fusion mechanism that integrates complementary information from RGB images and event data. In the broader Mamba literature, this places FocusMamba within a line of vision models that replace quadratic-complexity attention with Selective State Space Models for linear scalability and efficient long-range dependency modeling (Ibrahim et al., 11 Feb 2025).

1. Problem setting and design rationale

FocusMamba targets RGB-Event collaborative detection under the observation that background in images and non-event regions in event data are often processed unnecessarily, and that attention mechanisms in fusion stages may get distracted by background, missing key complementary regions (Yang et al., 4 Sep 2025). The paper further identifies a limitation in prior token sparsification methods: they employ a fixed number or threshold for token selection, which hinders the retention of informative tokens for samples with varying complexity. In the formulation given by the paper, this can produce under-pruning in simple scenes and over-pruning in complex scenes.

The architecture is therefore organized around two coupled objectives. The first is to reduce computation by collaboratively discarding low-information regions in both modalities. The second is to improve fusion by selectively integrating complementary features while suppressing background interference. This dual objective is realized through two modules: the Event-Guided Multimodal Sparsification strategy, abbreviated EGMS, and the Cross-Modality Focus Fusion module, abbreviated CMFF (Yang et al., 4 Sep 2025).

This design places event sensing in a control role rather than only a feature role. A plausible implication is that event activity is treated as a dynamic indicator of scene complexity, allowing the model to decide how much computation to allocate on a per-sample basis rather than by using a globally fixed sparsification policy.

2. Event-Guided Multimodal Sparsification

EGMS adaptively discards low-information tokens from both RGB and event modalities, per sample, and comprises a scoring module together with an Event-Guided Control Mechanism, or EGCM (Yang et al., 4 Sep 2025). For RGB images, the score of the $i$-th token is the $L_2$ activation of token features,

$${S_I}_i = \|X_i\|_2 = \sqrt{\sum_{j=1}^{C} X_{i,j}^2},$$

where $X_i$ is the feature of the $i$-th image token. For event data, scores are based on spatiotemporal event continuity. The timestamp accumulation is defined as

$S_E^{T}_{x, y} = \sum_{i,\, x_i = x,\, y_i = y} t_i,$

followed, after max-pooling, by a spatially weighted score

$S_E = \frac{\sum_{q \in \Omega} \left( \exp \left( -\frac{\|q - c\|^2}{2\sigma^2} \right) S_E^{T}_q \right)}{\sum_{q \in \Omega}\exp \left( -\frac{\|q - c\|^2}{2\sigma^2} \right)}.$

EGCM uses the event spatial ratio $r$, defined as the fraction of spatial positions triggered by events, to produce scene-adaptive scaling and thresholding. The scale factor is

$\text{Scale} = r^{\frac{1}{\rho}},$

and the normalized scores are

$$S_I = \text{softmax}\left( \frac{S_I}{\text{Scale}} \right), \qquad S_E = \text{softmax}\left( \frac{S_E}{\text{Scale}} \right).$$

The control factor is

$L_2$0

which determines the token selectivity masks

$L_2$1

These sparsification maps guide subsequent backbone layers to perform computation only in informative regions (Yang et al., 4 Sep 2025). The immediate significance is computational: the kept token ratio varies with observed scene complexity. The paper explicitly attributes to EGMS strong scene adaptivity, minimization of the loss of important content, and minimization of computational waste. This also clarifies a common misunderstanding about sparsification in multimodal detection: FocusMamba is not based on a fixed retention budget; its defining claim is sample-wise, modality-specific, event-guided selection.

3. Cross-Modality Focus Fusion

CMFF is the fusion stage that operates after sparsification. Its purpose is to capture and integrate complementary features from both modalities while ignoring background, using the differences between the sparsification maps $L_2$2 and $L_2$3 (Yang et al., 4 Sep 2025). It contains two components: Complementarity-Aware Enhancement, or CAE, and Focused Interlaced Mamba, or FI-Mamba.

CAE identifies degraded or underperforming regions in one modality by locating positions where the other modality preserves features and the first does not. For the image modality, the paper defines

$L_2$4

and then constructs an enhancement mask

$L_2$5

which yields enhanced image features

$L_2$6

The paper states that the equivalent operation is applied for event data. In effect, CAE amplifies regions where one modality can compensate for the other.

FI-Mamba then performs the actual joint modeling. It first gathers important tokens from both modalities at positions where at least one modality maintains information, using the union mask

$L_2$7

It then alternates tokens from both modalities to form a refined sequence, applies Bidi-Scan Mamba to capture intra- and cross-modal global dependencies efficiently, and scatters the enhanced tokens back to their original positions in the feature map. The features from both modalities are then summed and passed to a sparse MLP (Yang et al., 4 Sep 2025).

The union-based focus policy is central to the method. The ablation discussion states that focusing on the union, rather than the intersection or all tokens, is optimal. This suggests that FocusMamba is designed to privilege complementarity rather than redundancy: a token is preserved for fusion if either modality regards it as informative.

4. Mamba within the FocusMamba architecture

The Mamba component in FocusMamba appears in FI-Mamba as bidirectional selective state-space modeling over the interleaved multimodal token sequence (Yang et al., 4 Sep 2025). In the Mamba literature for vision, state-space models are characterized by linear scaling with input size, hardware-aware selective scanning, and the ability to model sequence dependencies with much less overhead than self-attention (Ibrahim et al., 11 Feb 2025). Surveys of visual Mamba further describe Selective Structured State Space Models as a mechanism for capturing long-range dependencies with linear computational complexity, in contrast to the quadratic complexity of Vision Transformers (Xu et al., 2024, Rahman et al., 2024).

Within that broader technical context, FocusMamba uses Mamba not as a generic replacement for all backbone computation, but as a focused fusion operator applied only after adaptive token selection. That distinction matters. The paper attributes computational savings not only to linear-complexity sequence modeling, but also to the fact that only informative and complementary areas are jointly processed (Yang et al., 4 Sep 2025). A plausible implication is that the efficiency gains arise from the interaction between sparsification and state-space fusion rather than from Mamba alone.

The paper’s description of FI-Mamba also aligns with survey characterizations of bidirectional scanning in visual Mamba models (Ibrahim et al., 11 Feb 2025). In FocusMamba, bidirectionality is used in a multimodal sequence in which image and event tokens are interleaved, so the state-space dynamics encode both intra-modal continuity and cross-modal context.

5. Empirical results and ablation evidence

The reported experiments are conducted on DSEC-Det and PKU-DAVIS-SOD (Yang et al., 4 Sep 2025). DSEC-Det is described as real-world driving with challenging variable lighting, 208k labels, and 8 classes. PKU-DAVIS-SOD is described as video with motion blur, low-light/static scenes, 1+ million box labels, and 3 classes.

Benchmark	FocusMamba result	Comparison stated in the paper
DSEC-Det, FocusMamba-B	mAP 34.6, FLOPs 60.8G	+2.0 mAP over baseline, -30% FLOPs; +4.2% mAP over SFNet using only 29% of its FLOPs
DSEC-Det, FocusMamba-S	mAP 32.3, FLOPs 35.9G	Small model
PKU-DAVIS-SOD, FocusMamba-B	mAP 32.7, FLOPs 30.2G	+0.8% mAP over SFNet with only 22% of its FLOPs

The paper states that FocusMamba outperforms state-of-the-art event-only, RGB-only, and other fusion models on DSEC-Det, and that FocusMamba-B achieves the best result on PKU-DAVIS-SOD (Yang et al., 4 Sep 2025). The qualitative analysis reports that FocusMamba retains more foreground and object tokens, especially in challenging illumination and motion scenes, and that it adapts token retention rate to scene complexity.

The ablation studies specify the role of each module. Adding EGMS reduces FLOPs by 33.1% and increases mAP by 0.8%. The scale and control factors are both described as crucial, with removal of either or both degrading scene adaptability and accuracy. Modality-specific scoring outperforms joint scoring and learned-score modules such as SAST and STCA. CAE and FI-Mamba each individually increase mAP, and together provide a 2.2% mAP boost. The paper also states that EGMS can be plugged into ViT-based frameworks, yielding significant FLOPs reduction and mAP gain (Yang et al., 4 Sep 2025).

These results delimit the paper’s central empirical claim. FocusMamba is presented not only as an efficient detector, but as an accuracy-efficiency trade-off method in which adaptive selection and focused fusion are mutually reinforcing.

6. Limitations, scope, and position in multimodal vision

The paper states a current limitation directly: the method may struggle in entirely static scenes, because the event spatial ratio is not informative for sparsification control, and it suggests integrating temporal information across frames to overcome this (Yang et al., 4 Sep 2025). This limitation is structurally tied to the method’s use of event activity as a control signal. It is therefore not merely a corner case in evaluation, but a boundary condition of the sparsification policy itself.

FocusMamba should also be distinguished from multimodal fusion approaches that apply dense processing uniformly to both modalities. Its defining idea is collaborative token sparsification before fusion, followed by Mamba-based processing over informative regions only. A common misconception would be to read it as a standard RGB-Event fusion network with a Mamba block added to the fusion stage. The paper’s module design and ablations indicate that the adaptive sparsification mechanism is co-equal with the state-space fusion mechanism, not ancillary to it.

In the broader Mamba ecosystem, FocusMamba exemplifies a specific application pattern that has become visible across vision research: Mamba is used where long-range dependency modeling and computational efficiency are both necessary, but the surrounding architecture is task-specific. Surveys of Mamba in vision emphasize linear computational complexity, selective state updates, and scan-based processing as the main architectural advantages (Rahman et al., 2024, Xu et al., 2024). FocusMamba instantiates those advantages in RGB-Event object detection by pairing bidirectional Mamba fusion with event-guided, scene-adaptive token selection (Yang et al., 4 Sep 2025).

The code is stated to be available at https://github.com/Zizzzzzzz/FocusMamba (Yang et al., 4 Sep 2025).