FogAct: A Real-World Benchmark for Foggy Action Recognition

Official Hugging Face repository for the paper:

Seeing Through Fog: Towards Fog-Invariant Action Recognition

Foggy conditions frequently appear in real-world scenarios such as surveillance, autonomous driving, and human-computer interaction. However, most existing action recognition methods are designed and evaluated under clear-weather assumptions. Under foggy conditions, reduced visibility, low contrast, and motion blur make it difficult for models to capture reliable action semantics.

To support research on robust action recognition under adverse weather, we introduce FogAct, the first real-world benchmark for foggy human action recognition with paired clean and foggy videos. We further propose FogNet, an end-to-end framework that learns fog-invariant action representations without relying on a separate defogging stage.

Highlights

• Real-world foggy action videos: FogAct is collected under real foggy conditions instead of being synthesized by atmospheric scattering models.
• Clean-foggy paired data: Each foggy video is paired with a clean counterpart captured by a synchronized stereo camera system.
• Large-scale action benchmark: FogAct contains 9,724 video clips, 55 action categories, and 10 diverse scenes.
• Multiple perspectives and fog levels: Actions are recorded from four perspectives and under light/dense fog conditions.
• Fog-invariant representation learning: FogNet learns robust action representations through Fog-Aware Selection, Mutual Enhancement, and Cross-Stream Alignment.

Dataset Overview

FogAct is collected using a stereo video acquisition system. The left camera records clean videos, while the right camera records the corresponding foggy videos generated by a professional fogging machine. The two cameras are synchronized to ensure semantic alignment between each clean-foggy video pair.

Item	Description
Dataset	FogAct
Task	Foggy action recognition
Number of videos	9,724
Number of action classes	55
Number of scenes	10
Resolution	1920 × 1080
Frame rate	25 FPS
Perspectives	Front, back, left, right
Fog levels	Light fog, dense fog
Data type	Paired clean and foggy videos

The dataset covers diverse real-world human actions, including pose-based actions, person-object interactions, and interpersonal interactions. Example actions include dribble basketball, listen to music, wave, mop floor, push, handshakes, and support somebody.

Method: FogNet

FogNet is an end-to-end foggy action recognition framework based on a two-stream CLIP architecture. Instead of first restoring foggy videos with a defogging model, FogNet directly learns semantic representations that are invariant to fog degradation.

FogNet contains three core components:

Fog-Aware Selection

Fog-Aware Selection uses global self-attention to identify semantically meaningful features from clean and foggy video embeddings, suppressing fog-induced degradation and emphasizing action-relevant cues.

Mutual Enhancement

Mutual Enhancement uses bidirectional cross-attention between clean and foggy streams. Clean videos guide the optimization of foggy representations, while foggy videos provide complementary information for robust feature learning.

Cross-Stream Alignment

Cross-Stream Alignment aligns enhanced clean and foggy video embeddings at the frame level. Since fog does not change the underlying action semantics, this module encourages temporal consistency between paired clean and foggy videos.

During training, FogNet uses both clean and foggy videos. During inference, only foggy videos are required.

Benchmark Results

Results on FogAct

Method	Top-1 Acc.	Top-5 Acc.
ActionCLIP	75.0	95.7
OST	83.2	98.9
PTTD + OST	85.4	98.8
FogNet	88.7	99.4

FogNet achieves the best performance on FogAct while avoiding the error accumulation caused by two-stage defogging-then-recognition pipelines.

Results on Simulated Foggy Datasets

Method	UCF-101	HMDB-51	Kinetics-100	Average
ActionCLIP	84.3	57.1	70.6	70.7
AIM	90.0	64.7	79.2	78.0
ATM	90.2	63.4	77.2	76.9
OST	92.4	69.1	75.4	79.0
FogNet	93.2	71.1	85.2	83.2

Usage

from datasets import load_dataset

dataset = load_dataset("YOUR_ORG/FogAct")
print(dataset)

Please replace YOUR_ORG/FogAct with the final Hugging Face repository name after release.

Intended Use

FogAct is intended for academic research on:

• Foggy action recognition
• Robust video understanding under adverse weather
• Fog-invariant representation learning
• Vision-language models for video recognition
• Defogging-free action recognition
• Domain adaptation between clean and degraded videos

Limitations

FogAct focuses on real-world foggy action recognition under controlled collection settings. Although the dataset includes multiple scenes, viewpoints, and fog intensities, it may not cover all possible weather conditions or deployment environments. Some subtle or short-duration actions remain challenging under dense fog, especially when different actions share similar motion patterns or visual appearances.

Ethical Considerations

The dataset was collected from volunteers who were informed about the intended use of the data. The data collection process followed ethical and legal requirements, and the annotations were manually reviewed.

FogAct should be used for academic and research purposes only. It should not be used for identity recognition, surveillance targeting, or applications that may violate privacy or harm individuals.

Citation

If you use FogAct or FogNet in your research, please cite our paper:

@article{liu2026seeing,
  title={Seeing Through Fog: Towards Fog-Invariant Action Recognition},
  author={Liu, Enqi and Pan, Liyuan and Gao, Zhi and Li, Lingzhi and Li, Qing},
  journal={arXiv preprint arXiv:2605.20645},
  year={2026}
}

Contact

For questions about the dataset or code, please contact:

• Enqi Liu: enqi.liu@bit.edu

Acknowledgement

This work is supported by the National Natural Science Foundation of China, the Fundamental Research Funds for the Central Universities, the BIT Special-Zone, and the Opening Project of the State Key Laboratory of General Artificial Intelligence.