Anytime Dense Prediction with Confidence Adaptivity

Authors: Zhuang Liu, Zhiqiu Xu, Hung-Ju Wang, Trevor Darrell, Evan Shelhamer

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate ADP-C on Cityscapes semantic segmentation and MPII human pose estimation: our method enables anytime inference without sacrificing accuracy while also reducing the total FLOPs of its base models by 44.4% and 59.1%.
Researcher Affiliation Collaboration Zhuang Liu1 Zhiqiu Xu1 Hung-Ju Wang1 Trevor Darrell1 Evan Shelhamer2 1University of California, Berkeley 2Adobe Research
Pseudocode No The paper describes its methods in prose and equations but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/liuzhuang13/anytime.
Open Datasets Yes The Cityscapes dataset (Cordts et al., 2016)... MPII Human Pose dataset (Andriluka et al., 2014)...
Dataset Splits Yes We train the models with the training set and report results on the validation set. ...on its validation set.
Hardware Specification Yes (specifically we measure computation time on a Linux machine with Intel Xeon Gold 5220R CPUs using 16 threads).
Software Dependencies No Our experiments are implemented using Py Torch (Paszke et al., 2019). No specific version number for PyTorch or other libraries is provided.
Experiment Setup Yes During training, multi-scale and flipping data augmentation is used, and the input cropping size is 512 1024. The model is trained for 484 epochs, with an initial learning rate of 0.01 and a polynomial schedule of power 0.9, a weight decay of 0.0005, a batch size of 12, optimized by SGD with 0.9 momentum.