Aligning Pretraining for Detection via Object-Level Contrastive Learning

Authors: Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, Stephen Lin

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, the proposed So Co achieves state-of-the-art transfer performance from Image Net to COCO.
Researcher Affiliation Industry Fangyun Wei Yue Gao Zhirong Wu Han Hu Stephen Lin Microsoft Research Asia {fawe, yuegao, wuzhiron, hanhu, stevelin}@microsoft.com
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/hologerry/SoCo.
Open Datasets Yes We adopt the widely used Image Net [1] which consists of 1.28 million images for self-supervised pretraining. COCO [42] and Pascal VOC [43] datasets are used for transfer learning.
Dataset Splits Yes We use the COCO train2017 set which contains 118k images with bounding box and instance segmentation annotations in 80 object categories. Transfer performance is evaluated on the COCO val2017 set.
Hardware Specification Yes The total batch size is set to 2048 over 16 Nvidia V100 GPUs.
Software Dependencies No The paper mentions software like 'Detectron2 [44] is used as the code base.' and 'We use MMDetection [49] as code base' but does not specify version numbers for these or any other software components.
Experiment Setup Yes We use a 100-epoch training schedule... We use the LARS optimizer [39] with a cosine decay learning rate schedule [40] and a warm-up period of 10 epochs. The base learning rate lrbase is set to 1.0... The weight decay is set to 1.0 e 5. The total batch size is set to 2048...