R-FCN++: Towards Accurate Region-Based Fully Convolutional Networks for Object Detection

Authors: Zeming Li, Yilun Chen, Gang Yu, Yangdong Deng

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our approach achieves state-of-the-art single-model results on both Pascal VOC and MS COCO object detection benchmarks, 87.3% on Pascal VOC 2012 test dataset and 42.3% on COCO 2015 testdev dataset. Our experiments use publicly available py-R-FCN code1. Following the setting of R-FCN, all experiments adopt atrous (Mallat 1999; Long, Shelhamer, and Darrell 2015; Chen et al. 2014) algorithm and online hard example mining (OHEM) (Shrivastava, Gupta, and Girshick 2016). Our models are initialized with standard Res Net-101 (He et al. 2016) or Res Net-50 which is pre-trained on the Image Net (Russakovsky et al. 2015) classification dataset.
Researcher Affiliation Collaboration Zeming Li,1 Yilun Chen,2 Gang Yu,2 Yangdong Deng1 1School of Software, Tsinghua University, {lizm15@mails.tsinghua.edu.cn, dengyd@tsinghua.edu.cn } 2 Megvii Inc. (Face++), {chenyilun, yugang}@megvii.com
Pseudocode No The paper describes the proposed methods using text and diagrams (Figure 2), but does not include any structured pseudocode or algorithm blocks.
Open Source Code No Code will be made publicly available. Our experiments use publicly available py-R-FCN code1. 1https://github.com/Orpine/py-R-FCN
Open Datasets Yes Our models are initialized with standard Res Net-101 (He et al. 2016) or Res Net-50 which is pre-trained on the Image Net (Russakovsky et al. 2015) classification dataset. We validate our method on Pascal VOC (Everingham et al. 2010; 2015) and MS COCO (Lin et al. 2014) dataset.
Dataset Splits Yes For the experiments on VOC, our algorithm is trained on Pascal VOC 2007+2012 trainval dataset, and tested on Pascal VOC 2007 test dataset. [...] For the COCO experiments, our algorithm is trained on COCO train set and test on minival set.
Hardware Specification Yes Only one Pascal TITAN GPU is used to train our model.
Software Dependencies No The paper mentions using 'py-R-FCN code' but does not specify any version numbers for this or other software dependencies, making it difficult to reproduce the exact software environment.
Experiment Setup Yes We use a weight decay of 0.0005 and a momentum of 0.9. The detection results are measured by mean Average Precision(m AP). For single scale training, the shorter edge of image is resized to 600 pixels. [...] Learning rate is set to 0.001 for first 80k iterations and 0.0001 for later 40k iterations. The iter size is set to 2. For the COCO experiments, our algorithm is trained on COCO train set and test on minival set. Learning rate is set to 0.0005 for first 1.2m iterations and 0.00005 for later 720k iterations. The iter size for COCO is set to 1.