Learning to Segment Object Candidates

Authors: Pedro O. O. Pinheiro, Ronan Collobert, Piotr Dollar

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We train on MS COCO [21] and evaluate the model on two object detection datasets, PASCAL VOC [7] and MS COCO.In this section, we evaluate the performance of our approach on the PASCAL VOC 2007 test set [7] and on the first 5000 images of the MS COCO 2014 validation set [21].
Researcher Affiliation Collaboration Pedro O. Pinheiro Ronan Collobert Piotr Doll ar pedro@opinheiro.com locronan@fb.com pdollar@fb.com Facebook AI Research Pedro O. Pinheiro is with the Idiap Research Institute in Martigny, Switzerland and Ecole Polytechnique F ed erale de Lausanne (EPFL) in Lausanne, Switzerland. This work was done during an internship at FAIR.
Pseudocode No The paper describes the network architecture and procedures in text and with diagrams (Figure 1), but does not include any structured pseudocode or algorithm blocks.
Open Source Code No All the experiments were conducted using Torch71. 1http://torch.ch. This link refers to the Torch7 framework used, not the specific source code for the methodology described in the paper. There is no explicit statement about releasing their code.
Open Datasets Yes We train on MS COCO [21] and evaluate the model on two object detection datasets, PASCAL VOC [7] and MS COCO. The references [21] and [7] provide citations to these publicly available datasets.
Dataset Splits Yes Our model is trained on the COCO training set which contains about 80,000 images and a total of nearly 500,000 segmented objects. and Design architecture and hyper-parameters were chosen using a subset of the MS COCO validation data [21] (non-overlapping with the data we used for evaluation). and We evaluate the performance of our approach on the PASCAL VOC 2007 test set [7] and on the first 5000 images of the MS COCO 2014 validation set [21].
Hardware Specification Yes Our model takes around 5 days to train on a Nvidia Tesla K40m.
Software Dependencies No All the experiments were conducted using Torch71. 1http://torch.ch. This mentions Torch7, but does not provide a specific version number for Torch or other software dependencies with their versions.
Experiment Setup Yes We considered a learning rate of .001. We trained our model using stochastic gradient descent with a batch size of 32 examples, momentum of .9, and weight decay of .00005. and Aside from the pre-trained VGG features, weights are initialized randomly from a uniform distribution. and Both of these layers are followed by Re LU non-linearity and a dropout [28] procedure with a rate of 0.5. and To binarize predicted masks we simply threshold the continuous output (using a threshold of .1 for PASCAL and .2 for COCO).