SAPA: Similarity-Aware Point Affiliation for Feature Upsampling

Authors: Hao Lu, Wenze Liu, Zixuan Ye, Hongtao Fu, Yuliang Liu, Zhiguo Cao

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate SAPA across a number of mainstream dense prediction tasks, for example: i) semantic segmentation: we test SAPA on several transformer-based segmentation baselines on the ADE20K dataset [6], such as Seg Former [7], Mask Former [8], and Mask2Former [9], improving the baselines by 1% 2.7% m Io U; ii) object detection: SAPA improves the performance of Faster R-CNN by 0.4% AP on MS COCO [10]; iii) monocular depth estimation: SAPA reduces the rmse metric of BTS [11] from 0.419 to 0.408 on NYU Depth V2 [12]; and iv) image matting: SAPA outperforms a strong A2U matting baseline [5] on the Adobe Composition-1k testing set [13] with a further 3.8% relative error reduction in the SAD metric.
Researcher Affiliation Academia Hao Lu Wenze Liu Zixuan Ye Hongtao Fu Yuliang Liu Zhiguo Cao School of Artificial Intelligence and Automation Huazhong University of Science and Technology Wuhan 430074, China {hlu,zgcao}@hust.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at: https://github.com/poppinace/sapa
Open Datasets Yes For semantic segmentation, we conduct experiments on the ADE20K dataset [6]... For object detection, we use the MS COCO [10] dataset... For depth estimation, we use NYU Depth V2 dataset [12]... For image matting, we train the model on the Adobe Image Matting dataset [13]...
Dataset Splits Yes For depth estimation, we use NYU Depth V2 dataset [12] and its default train/test split.
Hardware Specification Yes All our experiments are run on a server with 8 NVIDIA Ge Force RTX 3090 GPUs.
Software Dependencies No The paper mentions software like 'mmdetection' but does not specify version numbers for any key software components or libraries.
Experiment Setup Yes All training settings and implementation details are kept the same as the original papers. ... We set d = 32 and K = 5 in SAPA. ... For object detection... we follow the 1 (12 epochs) training configurations.