The common paradigm of CNN-based action recognition modelsis to simply use the average of the dense predictions from every frame. However, these dense predictions are inefficient since all frames are evenly utilized regardless of the existence of the action. In real-time action recognition applications where the input video is untrimmed, the dense prediction is even more severely inefficient. Instead of these dense predictions, we propose a lightweight CNN-based sampler that selects frames that are relevant to the action in the video. The proposed sampler is trained by a self-supervised learning method that mixes one irrelevant frame in an input video clip and finds that irrelevant frame. With the trained sampler, we select frames to be used as inputs to the action recognition model. Our sampler can be combined with any action recognition models and works independently of them. Through various experiments, we demonstrate that the proposed sampler is not only time and memory-efficient but also significantly increases performance in the action recognition benchmarks, including HMDB51 and UCF101. Finally, we applied the self-supervised sampler to the real-time surveillance system in the subway station, and improved the action recognition performance by 7% with 80% reduction in computation. Code and dataset are available in https://github.com/seominseok0429/SelSupSampler and https://aihub.or.kr/aidata/34124 .

Floods, exacerbated by climate change, necessitate timely and accurate situational awareness to support effective disaster response. While electro-optical (EO) satellite imagery has been widely employed for flood assessment, its utility is significantly limited under conditions such as cloud cover or nighttime. Synthetic Aperture Radar (SAR) provides consistent imaging regardless of weather or lighting conditions but it remains challenging for human analysts to interpret. To bridge this modality gap, we present diffusion-based SAR-to-EO image translation (DSE), a novel framework designed specifically for enhancing the interpretability of SAR imagery in flood scenarios. Unlike conventional GAN-based approaches, our DSE leverages the Brownian Bridge Diffusion Model to achieve stable and high-fidelity EO synthesis. Furthermore, it integrates a self-supervised SAR denoising module to effectively suppress SAR-specific speckle noise, thereby improving the quality of the translated outputs. Quantitative experiments on the SEN12-FLOOD dataset show that our method improves PSNR by 3.23 dB and SSIM by 0.10 over conventional SAR-to-EO baselines. Additionally, a user study with SAR experts revealed that flood segmentation performance using synthetic EO (SynEO) paired with SAR was nearly equivalent to using true EO–SAR pairs, with only a 0.0068 IoU difference. These results confirm the practicality of the DSE framework as an effective solution for EO image synthesis and flood interpretation in SAR-only environments.

Multi-task learning is a computationally efficient method to solve multiple tasks in one multi-task model, instead of multiple single-task models. MTL is expected to learn both diverse and shareable visual features from multiple datasets. However, MTL performances usually do not outperform single-task learning. Recent MTL methods tend to use heavy task-specific heads with large overheads to generate task-specific features. In this work, we (1) validate the efficacy of MTL in low-data conditions with early-exit architectures, and (2) propose a simple feature filtering module with minimal overheads to generate task-specific features. We assume that, in low-data conditions, the model cannot learn useful low-level features due to the limited amount of data. We empirically show that MTL can significantly improve performances in all tasks under low-data conditions. We further optimize the early-exit architecture by a sweep search on the optimal feature for each task. Furthermore, we propose a feature filtering module that selects features for each task. Using the optimized early-exit architecture with the feature filtering module, we improve the 15.937% in ImageNet and 4.847% in Places365 under the low-data condition where only 5% of the original datasets are available. Our method is empirically validated in various backbones and various MTL settings.

The Ensemble and mixture of expertise method is the most intuitive and simple way to improve performance in the field of recognition using convolutional neural networks (CNNs). However, Ensemble is difficult to apply in real-time operation applications because the amount of computational overhead and parameters increase in proportion to the number of models. In another, a mixture of expertise that extracts various expertise and combines it is cumbersome to apply because it requires a large change in the network. In this study, we propose Sequential Feature Filtering Classifier (FFC), a simple but effective classifier for CNNs. Using sequential LayerNorm and ReLU, FFC zeroes out low-activation units and preserves high-activation units. The sequential feature filtering process generates multiple features, which are transmitted to a shared classifier, yielding multiple outputs (multiple expertise). FFC can be applied to any CNN with a classifier and it significantly improves the performance with negligible overhead. In this study, the efficacy of FFC is validated extensively on various tasks—ImageNet-1K classification, MS COCO detection, Cityscapes segmentation, and HMDB51 action recognition. Moreover, it is empirically established that FFC can further improve performances using additional techniques, including attention modules. Code is available at https://github.com/seominseok0429/Sequential-Feature-Filtering-Classifier.