IAAF: Intersection-Aware Adaptive Fusion

• The paper presents IAAF as a lightweight feature fusion module that subtracts learned overlapping features between exposures to suppress ghosting artifacts.
• IAAF employs a compact CNN architecture with convolution layers and a residual block to explicitly estimate and remove redundant shared content, enhancing efficiency.
• Integrating IAAF within EfficienT-HDR improves HDR reconstruction in dynamic scenes by reducing ghosting and lowering computational costs for edge deployment.

Searching arXiv for the cited papers and adjacent work on HDR reconstruction and ghost suppression. Searching for ([2509.19779](/papers/2509.19779)) EfficienT-HDR and related HDR fusion papers. Intersection-Aware Adaptive Fusion (IAAF) is a feature-level fusion module introduced for multi-exposure HDR reconstruction in the EfficienT-HDR framework. Its stated purpose is to suppress ghosting artifacts caused by inter-exposure misalignment in dynamic scenes while remaining lightweight enough for edge deployment. In that formulation, IAAF replaces a more expensive spatial-attention-style fusion mechanism and treats exposure-specific feature maps as overlapping sets, explicitly estimating their shared component and subtracting it during fusion. The result is a ghost-suppression mechanism that is also a computational-efficiency mechanism within a transformer-based HDR pipeline (Huang et al., 24 Sep 2025).

1. Problem setting and motivation

In multi-exposure HDR reconstruction, three differently exposed LDR images are fused to recover details across shadows and highlights. The central difficulty addressed by IAAF is that dynamic scenes violate the static-scene assumption: moving objects or camera shake induce misalignment across exposures, and naive fusion of inconsistent regions produces ghosting artifacts, including doubled edges, blurry motion trails, and structural corruption. EfficienT-HDR identifies this failure mode, together with high computational cost, as the two main bottlenecks of Multi-Exposure Fusion (MEF) and HDR methods (Huang et al., 24 Sep 2025).

Within that setting, IAAF is introduced as the mechanism specifically responsible for handling motion-induced inconsistency. Its design objective is not merely generic multimodal or multi-branch fusion. Rather, it is targeted at the case where exposure branches contain overlapping but not perfectly aligned content, so that directly aggregating all features risks preserving duplicated or conflicting structures. The module is therefore framed around the notion of an estimated intersection between exposure features.

A common misconception is to treat IAAF as an attention reweighting block. In EfficienT-HDR, its role is more specific: it is an intersection-aware fusion operator intended to remove redundant or misaligned shared content before downstream reconstruction. The paper also presents it as a lightweight alternative to spatial attention, linking ghost suppression to deployment efficiency rather than treating them as separate design goals.

2. Architectural definition

Architecturally, IAAF is an intersection-aware, feature-level fusion module operating on exposure-specific feature maps. The inputs described in the model are $F_1$, $F_2$, and $F_3$, corresponding respectively to the over-exposed branch, the normal-exposed branch, and the under-exposed branch. The module is applied pairwise relative to the middle exposure $F_2$, specifically for the pairs $(F_1, F_2)$ and $(F_3, F_2)$ (Huang et al., 24 Sep 2025).

For each pair, the two feature maps are first concatenated. The concatenated tensor is then processed by a compact CNN consisting of one convolution layer, one residual block, and another convolution layer. This CNN predicts the shared component between the two inputs. The learned intersection is therefore not a hand-designed mask; it is a feature map estimated by a small convolutional subnetwork.

The output of IAAF is a fused feature representation in which ghosting is suppressed. In the main model, these IAAF-produced features are subsequently concatenated with original features and then combined with the middle-frame features for reconstruction. In the lite variant, only aligned features are concatenated with $F_2$, reducing the channel budget.

This architectural definition is significant because it shifts the fusion problem from weighting all spatial positions to estimating overlap explicitly. In effect, the module assumes that exposure branches contain both complementary information and a shared component whose direct duplication is undesirable under misalignment.

3. Fusion rule and mathematical interpretation

The core fusion rule given for IAAF is

$$F_{fused_1} \approx F_1 + F_2 - \text{Intersection}(F_1,F_2)$$

and

$$F_{fused_3} \approx F_3 + F_2 - \text{Intersection}(F_3,F_2).$$

A contextual formalization in the paper’s description is

$$\text{Intersection}(F_a, F_b) \approx \Phi([F_a, F_b]),$$

where $F_2$0 denotes the compact CNN composed of Conv, Residual Block, and Conv. The corresponding fused output is then expressed conceptually as

$F_2$1

The interpretation attached to these expressions is direct. The term $F_2$2 combines information from both exposures. The estimated intersection models what is redundant or common between them. Subtracting that shared component removes duplicated content and preserves complementary information. In the HDR setting, this is the mechanism by which the module is intended to reduce ghosting: mismatched or duplicated structures are less likely to survive the fusion stage when overlap is modeled explicitly rather than implicitly.

This does not amount to a fully elaborated set-theoretic formalism. The paper does not provide a more detailed objective for the intersection estimator, nor an explicit standalone intersection loss. The method is instead defined by the fusion rule itself and the small CNN that parameterizes the overlap estimate.

4. Placement within EfficienT-HDR

IAAF is located in the front end of the EfficienT-HDR pipeline. The sequence described for the framework is: input RGB LDR images at different exposures; conversion to YCbCr; feature extraction; IAAF-based fusion and ghost suppression; processing by the Context-Aware Vision Transformer (CA-ViT) backbone; use of Inverted Residual Embedding (IRE), Dynamic Tanh (DyT), and Enhanced Multi-Scale Dilated Convolution (E-MSDC); and final HDR reconstruction (Huang et al., 24 Sep 2025).

The relation between IAAF and YCbCr is functional rather than constitutive. YCbCr separates luminance from chrominance, and the paper states that IAAF benefits from this earlier conversion because HDR quality is most sensitive to luminance detail and exposure inconsistency. IAAF itself is not the color-space conversion module, but it operates on features prepared by that decomposition.

Its relation to CA-ViT is similarly staged. IAAF feeds the transformer with cleaner, better-aligned features, while CA-ViT models global dependencies via Multi-head Self-Attention and local context via the Local Context Extractor (LCE). In this formulation, IAAF reduces the burden on the transformer backbone by removing obvious redundancy and misalignment beforehand.

The distinction between the main and lite versions of EfficienT-HDR is also relevant to IAAF’s role. In the main model, IAAF-produced features are concatenated with original features and then fused with middle-frame features $F_2$3, yielding a feature budget of $F_2$4. In Ours-Lite, only aligned features are concatenated with $F_2$5, reducing the channel budget to $F_2$6. The module is therefore present in both variants, but the downstream usage of its outputs is altered to trade visual quality against efficiency.

5. Computational rationale and training context

A central rationale for IAAF is that it is lighter than spatial attention. The paper contrasts the two by stating that spatial attention requires large matrix operations and that its cost scales quadratically with the number of pixels or tokens, whereas IAAF’s convolutional intersection computation scales linearly with feature-map size. On that basis, IAAF functions simultaneously as a ghost-removal mechanism and as a computational-efficiency mechanism (Huang et al., 24 Sep 2025).

This efficiency claim is embedded in a broader lightweight design that also includes IRE, DyT, and E-MSDC. The aggregate result reported for the main model, relative to the HDR-Transformer baseline, is a reduction in FLOPS by approximately 67%, an inference speed increase of more than fivefold on CPU, and Jetson NX speeds of $F_2$7 ms/patch for the main model and $F_2$8 ms/patch for Ours-Lite. The paper does not attribute these gains exclusively to IAAF, and it is important not to collapse a system-level improvement into a single-module claim.

IAAF is trained as part of the full EfficienT-HDR system rather than through a dedicated auxiliary objective. The training configuration reported for the framework uses AdamW, a learning rate of $F_2$9, weight decay of $F_3$0, Cosine Annealing Warm Restarts, 100 epochs, batch size 14, training patches of $F_3$1, and the Joint Recon Perceptual Loss. One implication is that the intersection estimator is supervised only indirectly through the end-to-end reconstruction and perceptual objective, not through an explicitly described overlap-consistency term.

Another common misunderstanding is to assume that “adaptive” in IAAF denotes a learned scalar weighting rule analogous to attention logits or gating coefficients. In EfficienT-HDR, the adaptivity resides in the learned estimation of the shared component by the compact CNN and in the module’s use of exposure-pair-specific features, not in a separately specified weighting formula.

6. Empirical role, scope, and terminological ambiguity

The empirical evidence presented for IAAF’s role is primarily qualitative. On the Kalantari and Ramamoorthi dataset, challenging scenes with significant foreground motion are described as producing pronounced ghosting in traditional methods such as Sen12 and Kalantari17, whereas the proposed main model produces sharp object contours, much better suppression of ghosting, and better preservation of background structure. The light-weight version is also reported to suppress most ghosting, although with slightly reduced edge sharpness. This suggests that the alignment-aware fusion stage is a major contributor to dynamic-scene robustness within EfficienT-HDR (Huang et al., 24 Sep 2025).

At the same time, the scope of the term should be delimited carefully. “IAAF” is not a standardized acronym with a single cross-domain meaning. In a separate robotics paper, the same acronym is used for a contact-aware adaptive fusion strategy in diffusion-based manipulation, where it denotes contact gating plus classifier-free-guidance-style blending of vision and torque experts rather than intersection-aware feature fusion for HDR reconstruction (Lei et al., 1 Apr 2026). The acronym therefore has domain-specific meanings, and the HDR usage should not be conflated with the manipulation usage.

A further misconception is to interpret IAAF as a generic late-fusion framework. In EfficienT-HDR it is neither a general-purpose fusion abstraction nor a token-level attention block. It is a specific front-end module for pairwise exposure fusion centered on the equation “sum both branches, subtract the learned intersection.” Its technical identity is thus tied to ghost suppression under exposure misalignment, compact convolutional estimation of shared content, and integration into a lightweight transformer-based HDR pipeline.