Revisiting the Integration of Convolution and Attention for Vision Backbone

Authors: Lei Zhu, Xinjiang Wang, Wayne Zhang, Rynson Lau

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments on various vision tasks, we empirically verify the potential of the proposed integration scheme, named GLMix
Researcher Affiliation Collaboration Lei Zhu City University of Hong Kong ray.leizhu@outlook.com Xinjiang Wang Sensetime Research wangxinjiang@sensetime.com Wayne Zhang Sensetime Research wayne.zhang@sensetime.com Rynson Lau City University of Hong Kong Rynson.Lau@cityu.edu.hk
Pseudocode No The paper does not contain a block explicitly labeled as "Pseudocode" or "Algorithm", nor does it present structured steps in a pseudocode format.
Open Source Code No Code will be available at https://github.com/rayleizhu/GLMix.
Open Datasets Yes We conduct image classification experiments on the Image Net-1k dataset [13]... We evaluate the backbones for object detection and instance segmentation on COCO 2017 [31]... Our semantic segmentation experiments are conducted on the ADE20K dataset...
Dataset Splits No The paper refers to using "standard training recipes" and existing datasets like ImageNet-1k, COCO, and ADE20K, but it does not explicitly provide the specific validation dataset splits (percentages, counts, or explicit statements about using a predefined validation split from these datasets) within its own text.
Hardware Specification Yes following the same hardware (a single Tesla V100 32G GPU) and batch size (128) configurations used in Swin-Transformer [35].
Software Dependencies No The paper mentions using "the timm library [54]", "the MMDetection [4] toolbox", and "the MMSegmentation [10] toolbox" but does not specify the version numbers for these software components.
Experiment Setup Yes For the standard supervised training recipe, training details are in Table 8. When training with the advanced distillation recipe [26], we add an extra distillation head to the GLNet-4G/9G model and use the NFNet-F6 [2] to generate distillation targets; other training details are shown in Table 9.