acceptodds
Withdrawn from ICLR 2027

FREST: Frequency Representation Enhancement for Spatiotemporal Robustness in Infrared and Visible Video Fusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Compared with single-frame image fusion, infrared and visible video fusion better reflects real-world sensing scenarios and supports continuous scene perception. However, existing spatial-domain methods struggle to separate degradation interference from informative structures and distinguish abnormal temporal fluctuations from valid motion, resulting in limited robustness to complex degradations and compromised temporal coherence. To address these issues, we propose FREST, which leverages frequency representations to suppress interference from degradation and abnormal temporal fluctuations while preserving informative content and valid motion. Specifically, it comprises two components: degradation-robust spatial-frequency modeling and frequency-aware spatiotemporal modeling. On the one hand, to suppress degradation-sensitive responses while preserving scene structures, we develop a degradation-robust spatial-frequency modeling framework to exploit the distinct spectral characteristics of degradation interference and scene structures. Specifically, we develop an adaptive frequency enhancement module and a spatial-frequency collaborative modeling strategy for dynamic modulation and effective cross-modal interaction. On the other hand, given that abnormal flickering and valid motion exhibit distinct response patterns in the joint spatiotemporal frequency domain, we construct a frequency domain temporal consistency module to model spectral changes across consecutive frames, thereby reducing unstable temporal fluctuations while preserving meaningful motion. Extensive experiments demonstrate that FREST produces more temporally coherent fused videos than state-of-the-art methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
infrared and visible video fusion, deep learning, frequency-domain attention, temporal consistency

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3LxpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents FREST, a frequency-based framework for infrared and visible video fusion designed to improve robustness to image degradations and temporal inconsistencies.

The proposed approach consists of two main components: degradation-robust spatial-frequency modeling and frequency-aware spatiotemporal modeling. The first component uses adaptive frequency enhancement and spatial-frequency feature interaction to reduce degradation-related interference while retaining important visual details. It further incorporates frequency-guided cross-modal fusion to combine complementary information from infrared and visible videos. The second component employs 3D Fourier transforms to capture spatiotemporal frequency information, reducing unwanted temporal fluctuations while preserving meaningful motion.

Experiments on the M3SVD and HDO datasets show that FREST achieves state-of-the-art results in fusion quality and temporal consistency across various degradation scenarios, including blur, haze, rain, and infrared noise.

Critical strengths
  1. The paper presents a unified frequency-oriented framework for addressing both spatial degradations and temporal inconsistencies in infrared and visible video fusion. The overall motivation is reasonable and the proposed architecture is well structured.

  2. The authors introduce several dedicated modules for spatial-frequency enhancement, cross-modal fusion, and temporal consistency modeling. The experimental results across multiple degradation settings, together with the ablation studies, support the effectiveness of the proposed framework.

Critical weaknesses
  1. Unclear core technical contribution. Although the paper addresses spatial degradations and temporal inconsistencies in infrared-visible video fusion, it is not sufficiently clear which specific limitations of existing methods the proposed framework aims to overcome. Furthermore, the framework incorporates multiple frequency-based modules, but the key technical insights and distinctions from existing approaches are not clearly highlighted.

  2. Insufficient justification for the architectural design. The framework introduces multiple specialized modules, but the necessity of each component and the advantages over simpler alternatives are not fully established.

Questions for authors

Could the authors further clarify the key technical differences between the proposed frequency-based modules and existing spatial-frequency modeling approaches? In particular, what are the main methodological contributions beyond combining established techniques?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 3Lxp

R1: We would like to clarify that existing methods primarily address degradations by learning a hard mapping from degraded inputs to clean representations. However, degradation-induced variations and genuine scene dynamics are often spatially entangled (Fig. 1 in the original manuscript). These methods lack an explicit mechanism to distinguish abnormal degradation patterns from valid scene motion, which limits their ability to handle complex degradations while preserving meaningful dynamic information. The two most closely related works are VideoFusion and FTPFusion. VideoFusion employs bi-temporal co-attention to aggregate cross-frame information, while FTPFusion decomposes features into high- and low-frequency components to enhance temporal feature representations and robustness. Although FTPFusion incorporates frequency-domain processing, it primarily uses frequency decomposition for feature modeling rather than explicitly exploiting the spectral differences between degradation-induced disturbances and genuine scene dynamics.

Our key insight is that frequency-domain representations provide complementary cues for distinguishing these ambiguous signals (Fig. 1 in the original manuscript). Based on this insight, we develop two specialized modules to address spatial and temporal degradations from different perspectives. Specifically, SFCEM jointly models amplitude and phase information to selectively enhance informative spatial structures while reducing degradation-related interference. FTCM, rather than directly aggregating temporal features, performs joint spatiotemporal frequency modeling to capture the spectral characteristics of cross-frame variations, enabling the network to better preserve genuine scene dynamics while suppressing degradation-induced fluctuations.

R2: Thanks for this suggestion. Actually, our ablation studies have already demonstrated the effectiveness of the proposed modules, as removing any component leads to performance degradation. From a design perspective, FTCM exploits the distinct spectral responses of temporal degradations and valid scene dynamics to suppress degradation-induced fluctuations while preserving meaningful motion information. In contrast, SFCEM leverages frequency-domain characteristics to selectively enhance informative spatial structures and suppress degradation-related interference, thereby improving the preservation of complementary infrared and visible details. To further demonstrate these design necessity experimentally, we separately replace all six FTCM blocks and all ten SFCEM blocks with computationally comparable 3D convolutions while keeping the remaining architecture unchanged. The quantitative results on M3SVD and HDO are presented in Tables 1a and 1b.

Table 1a. Fusion quality and temporal consistency for the M3SVD dataset.

MethodMI ↑SSIM ↑flowD ↓BiSWE ↓MS2R ↓feaCD ↓
FREST w/ Conv3D replacing FTCM3.37240.62833.58633.75650.14471.7788
FREST w/ Conv3D replacing SFCEM2.67000.61086.03966.48770.14711.8630
FREST (ours)3.39200.63162.81682.83820.13981.6187

Table 1b. Fusion quality and temporal consistency for the HDO dataset.

MethodMI ↑SSIM ↑flowD ↓BiSWE ↓MS2R ↓feaCD ↓
FREST w/ Conv3D replacing FTCM2.75380.503210.82865.24120.60031.4845
FREST w/ Conv3D replacing SFCEM2.17810.475513.090010.23730.61611.7031
FREST (ours)2.77700.512010.51435.17530.58041.4445

As shown in Tables 1a and 1b, replacing FTCM or SFCEM with conventional 3D convolutions consistently degrades fusion quality and temporal consistency on both datasets, indicating that simply relying on conventional spatiotemporal convolutions is insufficient. These results further validate the necessity of the proposed frequency-domain modules for effectively modeling spatial and temporal information in infrared and visible video fusion.

Thank you for the response. My main concern remains novelty: the decoder and losses follow VideoFusion, FSAS/DFFN follow FFTformer, and FTCM closely resembles a 3D extension of FFC's spectral transform. The claimed "explicit" exploitation of spectral differences is not reflected in any specific mechanism, so the work reads as a combination of existing components.

Reviewer_hrzzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FREST for infrared and visible video fusion under degradations. The method introduces spatial-frequency enhancement and frequency-guided cross-modal fusion to preserve useful information from degraded inputs. A Frequency-Domain Temporal Consistency Module (FTCM) applies 3D Fourier transforms to model temporal variations. Experiments on M3SVD and HDO show improvements in fusion quality and temporal consistency over several image and video fusion methods.

Critical strengths
  1. The problem is meaningful. Handling source degradation and temporal instability together is relevant to practical infrared-visible video fusion.
  2. The frequency-domain temporal modeling is interesting. Removing FTCM clearly worsens flowD and BiSWE in the ablation, supporting the usefulness of temporal feature processing.
  3. The experiments include two datasets and several degradation settings. The results on HDO provide some evidence of generalization beyond the training dataset.
Critical weaknesses
  1. Figure 1 considers the same blur across scenes and a global brightness fluctuation. Wether these examples sufficient to fully support the broader argument that “degradation” can be distinguished from “valid scene content” through spectral response, particularly in cases involving local flickering, camera motion, or the movement of rain?
  2. For experiments, removing FTCM shows that the module helps, but does not establish the advantage of frequency-domain processing. A comparison with spatial-domain 3D convolution under a similar parameter budget is needed.
  3. The method adopts the decoder and training losses follow VideoFusion. However, VideoFusion is not included in the comparison.
  4. As can be seen from Table 3, the computational cost is very high, which raises concerns regarding its practical utility.
  5. Current image-level fusion methods are often evaluated through downstream tasks. Can the proposed method improve downstream performance within the heterogeneous degradations? Please provide quantitative experimental results.
Questions for authors

Refer to Weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer hrzz (1/2)

R1: Figure 1 in the original manuscript shows that identical blur or brightness fluctuations produce similar spectral responses across scenes. To examine more complex conditions, we simulate camera motion across 20 scenes with localized irregular flicker or moving rain streaks. Table 1 reports the median overlap between degradation-induced and motion-induced frame-difference energy distributions in the image and frequency domains, where higher values indicate greater similarity.

Table 1. Comparison of energy distribution overlap.

ComparisonImage DomainFrequency Domain
Local irregular flicker and camera motion0.6990.036
Moving rain streaks and camera motion0.4770.269

The results show that degradation and genuine motion can overlap spatially while exhibiting substantially different frequency distributions, supporting the use of spectral cues to distinguish degradation-induced disturbances from valid scene dynamics under the tested conditions.

R2: We replace all six FTCM blocks with parameter-matched 3D convolutions while keeping the remaining architecture unchanged. Both models use identical inputs, with 10.887 M parameters and 168.486 G convolutional MACs. As shown in Tables 2a and 2b, FREST improves all six metrics on both datasets. These results support the effectiveness of joint spatiotemporal frequency modeling over conventional convolutions.

Table 2a. FTCM and matched Conv3D on the M3SVD dataset.

ModelMI ↑SSIM ↑flowD ↓BiSWE ↓MS2R ↓feaCD ↓
FREST w/ Conv3D replacing FTCM3.37240.62833.58633.75650.14471.7788
FREST (ours)3.39200.63162.81682.83820.13981.6187

Table 2b. FTCM and matched Conv3D on the HDO dataset.

ModelMI ↑SSIM ↑flowD ↓BiSWE ↓MS2R ↓feaCD ↓
FREST w/ Conv3D replacing FTCM2.75380.503210.82865.24120.60031.4845
FREST (ours)2.77700.512010.51435.17530.58041.4445

R3: Unlike VideoFusion, which relies on bi-temporal co-attention for cross-frame aggregation, FREST employs frequency-domain modeling to distinguish valid scene dynamics from degradation-induced disturbances. Comparisons on M3SVD and HDO under four conditions (Tables 3a and 3b) show that VideoFusion achieves lower flowD and BiSWE in some cases, but its lower MI and SSIM and observed detail loss suggest a possible trade-off between temporal smoothness and spatial fidelity. In contrast, FREST achieves superior other metrics on both datasets, demonstrating a better balance between fusion quality and temporal consistency.

For comparison, FREST directly processes degraded inputs without any preprocessing. As VideoFusion can only handle several degradations, such as visible blur and infrared noise, we apply Turtle preprocessing only to additional degradations beyond its handling capabilities, namely VI haze and rain.

Table 3a. Fusion quality and temporal consistency on the M3SVD dataset.

MethodMI ↑SSIM ↑flowD ↓BiSWE ↓MS2R ↓feaCD ↓
VideoFusion2.87680.62002.07802.01630.15341.7889
FREST (ours)3.39200.63162.81682.83820.13981.6187

Table 3b. Fusion quality and temporal consistency on the HDO dataset.

MethodMI ↑SSIM ↑flowD ↓BiSWE ↓MS2R ↓feaCD ↓
VideoFusion1.87900.45707.82405.28030.63391.7009
FREST (ours)2.77700.512010.51435.17530.58041.4445
AuthorsOpenReview ↗

Response to Reviewer hrzz (2/2)

R4: We would like to clarify that video fusion generally requires more computation than image fusion due to additional temporal modeling. We acknowledge that FREST incurs higher computational overhead than lightweight methods such as MAVFusion. Specifically, FREST requires 0.255 s per frame, approximately 4.40× the inference time of MAVFusion (0.058 s). However, at a resolution of 640×480640×480, FREST requires 963.097 G FLOPs per frame, which is 11.82% and 78.70% lower than TemCoCo and UniVF, respectively. Compared with MAVFusion, FREST requires 2.91× FLOPs but reduces average BiSWE by 27.83% on M3SVD and 19.60% on HDO, while achieving the lowest MS2R and feaCD across all eight dataset-condition combinations. These improvements demonstrate the trade-off between computational efficiency and fusion quality under heterogeneous degradations. Regarding practical deployment, none of the compared video fusion methods including ours achieves the 30 FPS real-time requirement under the reported settings. We have clarified this limitation in the revised manuscript. Future work will focus on lightweight temporal modeling and efficient frequency-domain operations to reduce inference latency.

R5: We additionally evaluate downstream single-object tracking on degraded infrared (with stripe noise) and visible (with blur, haze, rain) sequences from NOT-156 [r1], using a frozen YOLO11n detector and ByteTrack tracker without fine-tuning (Section 4.4 of the revised manuscript). Despite requiring no external restoration, FREST achieves the best performance across all four tracking metrics, demonstrating its effectiveness for downstream tracking.

Table 4. Downstream tracking performance on the NOT-156 dataset.

Fusion MethodAUC ↑[email protected] ↑[email protected] ↑DP@20 ↑
LRRNet0.62960.81560.47090.8029
CrossFuse0.67220.88040.48160.8744
GIFNet0.59860.77760.45360.7809
FreeFusion0.59630.76550.45160.7589
TemCoCo0.40610.52440.30130.5137
UniVF0.62840.79690.52100.7983
FTPFusion0.73330.94120.61720.9325
VideoFusion0.66980.85500.53640.8624
MAVFusion0.67850.87040.53970.8704
FREST (ours)0.74720.94590.66200.9472

[r1] Sun C, Wang X, Fan S, et al. NOT-156: Night object tracking using low-light and thermal infrared: From multimodal common-aperture camera to benchmark datasets[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 1-11.

Reviewer_YvBFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates two key challenges in infrared and visible video fusion: interference from complex degradations and the impact of abnormal inter-frame fluctuations on temporal consistency. To address these challenges, the authors propose FREST, a multi-scale video fusion framework based on frequency representations. The proposed method employs a Frequency Enhancement Block (FreqEn) to suppress degradation interference while preserving informative details, a Frequency Modality-Guided Fusion (FMGF) module to integrate complementary infrared and visible information, and a Frequency-Domain Temporal Consistency Module (FTCM) to improve temporal coherence while preserving meaningful motion.

Experiments are conducted on the M3SVD and HDO datasets, covering relatively clean inputs and various degradation scenarios involving noise, blur, haze, and rain. The results demonstrate that FREST outperforms the compared methods in most evaluation settings, achieving competitive performance in fusion quality, degradation robustness, and temporal consistency.

Critical strengths
  1. This paper proposes a unified frequency-domain modeling framework to address complex degradations and temporal instability in infrared and visible video fusion. Specifically, FreqEn suppresses degradation interference while preserving informative details, FMGF enhances cross-modal interactions, and FTCM improves temporal coherence. The overall design is well motivated by the identified challenges, and the proposed components complement each other to jointly improve fusion quality, degradation robustness, and temporal stability.

  2. The proposed method is evaluated on two datasets, M3SVD and HDO, covering relatively clean inputs and various compound degradation settings, with comparisons against several representative image and video fusion methods. The experimental results show that FREST achieves favorable fusion quality and temporal consistency metrics in most evaluation settings. In addition, the ablation studies validate the effectiveness of the main components, while the qualitative comparisons and computational efficiency analysis provide further evidence for evaluating the proposed approach.

Critical weaknesses
  1. The paper provides comparisons with several representative image and video fusion methods, and the overall experimental coverage is reasonable. However, VideoFusion serves as an important basis for parts of the decoder design and the adopted loss functions, while FTPFusion is closely related to the proposed frequency-aware temporal modeling. The absence of direct comparisons with these methods makes it difficult to assess the extent of improvement over the most relevant prior work. More importantly, the current ablation studies primarily validate the effectiveness of the proposed components by removing them, which is insufficient to demonstrate their distinctive technical value or establish that they cannot be readily replaced by conventional alternatives. Therefore, I suggest conducting replacement experiments that compare the proposed frequency-domain modules with conventional spatial-domain or temporal-domain counterparts under the same network backbone and comparable computational budgets.

  2. The paper evaluates fusion quality and temporal consistency using MI, SSIM, flowD, and BiSWE, achieving favorable results. However, these metrics are insufficient to directly verify whether FTCM can effectively suppress abnormal flickering while accurately preserving the trajectories, boundaries, and details of moving objects. I suggest including controlled flickering experiments together with motion fidelity evaluations to further validate the effectiveness of the proposed frequency-aware spatiotemporal modeling.

  3. In Table 1, the complete FREST achieves an MI of 3.083 on M3SVD, whereas Table 2 reports MI values ranging from 2.959 to 3.617 under different degradation settings. The paper does not clearly explain the differences in evaluation conditions and data configurations between these two sets of experiments, which affects the interpretability and reproducibility of the reported results. I suggest that the authors explicitly clarify the degradation settings, test data, and evaluation protocol used in the ablation studies.

  4. Table 3 shows that FREST requires 0.255 seconds for inference, substantially longer than the 0.058 seconds reported for MAVFusion, corresponding to approximately 4.4 times the inference time. Although FREST achieves favorable results in fusion quality and temporal consistency, its relatively high computational cost may limit its applicability to real-time video fusion scenarios. The paper does not sufficiently analyze the trade-off between performance improvements and additional computational overhead, nor does it adequately discuss the applicability and deployment limitations of the proposed method in latency-sensitive scenarios such as real-time video fusion.

Questions for authors

Please refer to the concerns and suggestions raised in the Weaknesses section.

2 comments
AuthorsOpenReview ↗

Response to Reviewer YvBF (1/2)

R1: Thank you for the suggestion. We have added comparisons with VideoFusion and FTPFusion on infrared (IR) and visible (VI) videos under normal and degraded conditions, including IR noise paired with VI blur, haze, and rain. As shown in Tables 1a and 1b, VideoFusion achieves lower flowD and BiSWE in some cases, due to smoother temporal representations from cross-frame aggregation. However, its lower MI and SSIM, together with the observed loss of fine details, suggest that such smoothness may compromise spatial information. In contrast, FREST suppresses degradation-induced fluctuations while preserving meaningful dynamics and multimodal details. Its superior MI, SSIM, MS2R, and feaCD on both datasets indicate a better balance between fusion quality and temporal consistency.

For comparison, FREST directly processes degraded inputs without any preprocessing. Since FTPFusion lacks explicit degradation removal capabilities, we apply pretrained restoration models (MDIVDNet and Turtle) to its inputs under all degraded conditions. In contrast, VideoFusion can handle certain degradations, such as VI blur and IR noise. Therefore, we apply Turtle preprocessing only to additional degradations beyond its handling capabilities, namely VI haze and rain.

Table 1a. Fusion quality and temporal consistency for the M3SVD dataset.

MethodMI ↑SSIM ↑flowD ↓BiSWE ↓MS2R ↓feaCD ↓
VideoFusion2.87680.62002.07802.01630.15341.7889
FTPFusion3.01430.62302.84532.97930.14171.7316
FREST w/ Conv3D replacing FTCM3.37240.62833.58633.75650.14471.7788
FREST w/ Conv3D replacing SFCEM2.67000.61086.03966.48770.14711.8630
FREST (ours)3.39200.63162.81682.83820.13981.6187

Table 1b. Fusion quality and temporal consistency for the HDO dataset.

MethodMI↑SSIM↑flowD↓BiSWE↓MS2R↓feaCD↓
VideoFusion1.87900.45707.82405.28030.63391.7009
FTPFusion2.49450.507510.27485.81550.60891.5229
FREST w/ Conv3D replacing FTCM2.75380.503210.82865.24120.60031.4845
FREST w/ Conv3D replacing SFCEM2.17810.475513.090010.23730.61611.7031
FREST (ours)2.77700.512010.51435.17530.58041.4445

We further replace FTCM and SFCEM separately with parameter-matched 3D convolutions while preserving the remaining architecture. Both replacements degrade fusion quality and temporal consistency on both datasets, supporting the necessity of the proposed frequency-domain modules over conventional convolutions.

R2. We construct 10-frame M3SVD clips by shifting a 96×9696\times96 crop one pixel per frame and rendering a 20×2020\times20 textured target moving diagonally along a known trajectory in the visible input only. FREST and its parameter-matched Conv3D variant process identical clips with and without global brightness flicker (±0.10\pm 0.10 on a [0,1][0,1] scale). The metrics are defined as follows: in-phase flicker loss measures the retained fraction of injected flicker amplitude in the fused background; target-center error is the mean distance between the subpixel template-matched center and ground truth; boundary recall is the fraction of the true target boundary within 1.51.5 pixels of a detected edge; and texture correlation is the Pearson correlation between fused luminance and rendered texture inside the target. Table 2 shows that FREST outperformed the Conv3D variant across all metrics, reducing residual flicker by 16.89 percentage points, localizing the target with a 0.0985-pixel mean error, and achieving higher boundary recall and texture correlation.

Table 2. Quantitative comparison of residual flicker and target fidelity under alternating global flicker.

ModelIn‑phase flicker loss ↓Target‑center error (px) ↓Boundary recall ↑Texture correlation ↑
FREST w/ Conv3D replacing FTCM20.47%0.10580.92240.9865
FREST (ours)3.58%0.09850.94120.9997

R3: Thank you for pointing this out. The original Tables 1 and 2 evaluated different frame counts (100 vs. 150) sampled from the same test sequences under identical degradation types, severity levels, model weights, configurations, and metric procedures. Thus, the reported MI differences stemmed solely from evaluated frame subset variations rather than protocol inconsistencies. To ensure consistency and fair comparison with existing methods, we have standardized the evaluation to 200 frames and recomputed all quantitative results in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer YvBF (2/2)

R4: We would like to clarify that video fusion generally requires more computation than image fusion due to additional temporal modeling. We acknowledge that FREST incurs higher computational overhead than lightweight methods such as MAVFusion. Specifically, FREST requires 0.255 s per frame, approximately 4.40× the inference time of MAVFusion (0.058 s). However, at a resolution of 640×480640\times480, FREST requires 963.097 G FLOPs per frame, which is 11.82% and 78.70% lower than TemCoCo and UniVF, respectively. Compared with MAVFusion, FREST requires 2.91× FLOPs but reduces average BiSWE by 27.83% on M3SVD and 19.60% on HDO, while achieving the lowest MS2R and feaCD across all eight dataset-condition combinations. These improvements demonstrate the trade-off between computational efficiency and fusion quality under heterogeneous degradations. Regarding practical deployment, none of the compared video fusion methods including ours achieves the 30 FPS real-time requirement under the reported settings. We have clarified this limitation in the revised manuscript. Future work will focus on lightweight temporal modeling and efficient frequency-domain operations to reduce inference latency.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers: Summary of Manuscript Revisions

We thank all reviewers for their constructive comments. We have conducted additional experiments and clarified the technical contributions and limitations of FREST. The main revisions are summarized below.

Reviewer YvBF

  • Added comparisons with VideoFusion and FTPFusion under various degradation conditions.
  • Evaluated flicker suppression and motion preservation using controlled experiments.
  • Standardized the evaluation protocol to 200 frames and clarified computational costs and limitations.

Reviewer hrzz

  • Analyzed spectral differences between degradation-induced disturbances and genuine motion.
  • Conducted controlled comparisons between FTCM and parameter-matched 3D convolutions.
  • Added downstream object tracking experiments on NOT-156.
  • Extended comparisons with VideoFusion and clarified computational efficiency.

Reviewer 3Lxp

  • Clarified the novelty of FREST compared with existing video fusion methods.
  • Added Conv3D replacement experiments to validate the necessity of both modules.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.