This is the official implementation of our paper: "Deep Nets with Subsampling Layers Unwittingly Discard Useful Activations at Test-Time" accepted in ECCV 2024.
Subsampling layers play a crucial role in deep nets by discarding a portion of an activation map to reduce its spatial dimensions. This encourages the deep net to learn higher-level representations. Contrary to this motivation, we hypothesize that the discarded activations are useful and can be incorporated on the fly to improve models' prediction. To validate our hypothesis, we propose a search and aggregate method to find useful activation maps to be used at test-time. We applied our approach to the task of image classification and semantic segmentation. Extensive experiments over nine different architectures on ImageNet, CityScapes, and ADE20K show that our method consistently improves model test-time performance. Additionally, it complements existing test-time augmentation techniques to provide further performance gains.
You can set up the environment using the provided script.
bash scripts/tools/create_env.shPlease follow the instructions in Datasets Preparation.
- For a simple demo of how our modified subsampling layer work, please run the following code:
python3 test_toy.pybash scripts/cls/test/without_tta.shbash scripts/cls/test/with_tta.sh- This work is licensed under the Apache-2.0 license.
- Our project also involves the following assets from other research or projects.
@inproceedings{yangdeep,
title={Deep Nets with Subsampling Layers Unwittingly Discard Useful Activations at Test-Time},
author={Yang, Chiao-An, Liu, Ziwei, and Yeh, Raymond A},
booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
year={2024}
}Please contact Chiao-An Yang [[email protected]] if you have any questions.