SasWOT: Real-Time Semantic Segmentation Architecture Search WithOut Training

Authors: Chendi Zhu, Lujun Li, Yuli Wu, Zhengxing Sun

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on Cityscapes and Cam Vid datasets demonstrate that Sas WOT achieves superior trade-off between accuracy and speed over several state-of-the-art techniques.
Researcher Affiliation Academia 1State Key Laboratory for Novel Software Technology, Nanjing University 2The Hong Kong University of Science and Technology 3Institute of Imaging and Computer Vision, RWTH Aachen University, Aachen, Germany
Pseudocode Yes Algorithm 1: Evolution Search for Sas WOT Proxy
Open Source Code No The paper does not contain an explicit statement indicating that the source code for their methodology is publicly available, nor does it provide a link to a code repository.
Open Datasets Yes In this section, we first evaluate the ranking performance of the searched proxies in the semantic segmentation benchmark, and then we use the searched proxies to perform a training-free segmenter search in Cityscapes (Cordts et al. 2016) and Cam Vid (Brostow et al. 2008). The Trans NAS-Bench101 (Duan et al. 2021) dataset was used as our segmentation benchmark to evaluate the performance of the searched proxies.
Dataset Splits Yes Cam Vid is another street scene dataset... It contains 701 images in total, where 367 for training, 101 for validation and 233 for testing.
Hardware Specification No The paper mentions that the segmenter search can be 'completed on a single GPU within 2 hours' but does not provide any specific details about the GPU model (e.g., NVIDIA A100, RTX 3090) or other hardware specifications (CPU, RAM, etc.) used for the experiments.
Software Dependencies No The paper mentions using an 'SGD optimizer' and a 'deeplabv3+ model with a backbone of resnet101' but does not specify software dependencies with version numbers, such as Python version, deep learning framework (e.g., PyTorch, TensorFlow) version, or CUDA version.
Experiment Setup Yes For Cityscapes: 'For this task, we used an SGD optimizer with an initial learning rate of 0.015 and an exponential learning rate decay. We trained the searched architectures for 800 epochs'. For Cam Vid: 'We trained the searched architecture 80 epochs using the SGD optimizer, with an initial learning rate of 0.01 and exponential learning rate decay'.