@article{Iandola:2014tj,
author = {Iandola, F and Moskewicz, M and Karayev, S and Girshick, Ross B and Darrell, Trevor and Keutzer, K},
title = {{DenseNet: Implementing Efficient ConvNet Descriptor Pyramids}},
journal = {ArXiv e-prints},
year = {2014},
volume = {cs.CV},
month = apr,
annote = {essentially, a cheaper method to computing features on region proposals, by computing feature map of whole images at multiple scales, and cropping them to regiona proposals.


Actually such method dates back to older models such as DPM. Maybe because CNN features are global, not local (like HOG), and thus this method is not implemented before, as it's not accurate (since CNN has pooling, strides, etc;), compared to sliding window feature in HOG.

It's not elegant, such as aspect ratio, etc. Later on much better methods (Faster RCNN, etc.) are used. People find that a single scale might be sufficient.},
keywords = {deep learning},
read = {Yes},
rating = {2},
date-added = {2017-04-24T15:21:02GMT},
date-modified = {2017-04-24T15:32:26GMT},
url = {http://arxiv.org/abs/1404.1869},
local-url = {file://localhost/Users/yimengzh/Documents/Papers3_revised/Library.papers3/Articles/2014/Iandola/arXiv%202014%20Iandola.pdf},
file = {{arXiv 2014 Iandola.pdf:/Users/yimengzh/Documents/Papers3_revised/Library.papers3/Articles/2014/Iandola/arXiv 2014 Iandola.pdf:application/pdf}},
uri = {\url{papers3://publication/uuid/744F7A4F-AB1A-43F1-8F08-25A025987E62}}
}

{"_id":"NeZEsqn5ttvGgEs6Y","bibbaseid":"iandola-moskewicz-karayev-girshick-darrell-keutzer-densenetimplementingefficientconvnetdescriptorpyramids-2014","downloads":0,"creationDate":"2017-06-08T12:53:07.153Z","title":"DenseNet: Implementing Efficient ConvNet Descriptor Pyramids","author_short":["Iandola, F","Moskewicz, M","Karayev, S","Girshick, R. B","Darrell, T.","Keutzer, K"],"year":2014,"bibtype":"article","biburl":"https://leelabcnbc.github.io/lab-wiki/reference_library/computer_vision/object_detection_localization/bib.bib","bibdata":{"bibtype":"article","type":"article","author":[{"propositions":[],"lastnames":["Iandola"],"firstnames":["F"],"suffixes":[]},{"propositions":[],"lastnames":["Moskewicz"],"firstnames":["M"],"suffixes":[]},{"propositions":[],"lastnames":["Karayev"],"firstnames":["S"],"suffixes":[]},{"propositions":[],"lastnames":["Girshick"],"firstnames":["Ross","B"],"suffixes":[]},{"propositions":[],"lastnames":["Darrell"],"firstnames":["Trevor"],"suffixes":[]},{"propositions":[],"lastnames":["Keutzer"],"firstnames":["K"],"suffixes":[]}],"title":"DenseNet: Implementing Efficient ConvNet Descriptor Pyramids","journal":"ArXiv e-prints","year":"2014","volume":"cs.CV","month":"April","annote":"essentially, a cheaper method to computing features on region proposals, by computing feature map of whole images at multiple scales, and cropping them to regiona proposals. Actually such method dates back to older models such as DPM. Maybe because CNN features are global, not local (like HOG), and thus this method is not implemented before, as it's not accurate (since CNN has pooling, strides, etc;), compared to sliding window feature in HOG. It's not elegant, such as aspect ratio, etc. Later on much better methods (Faster RCNN, etc.) are used. People find that a single scale might be sufficient.","keywords":"deep learning","read":"Yes","rating":"2","date-added":"2017-04-24T15:21:02GMT","date-modified":"2017-04-24T15:32:26GMT","url":"http://arxiv.org/abs/1404.1869","local-url":"file://localhost/Users/yimengzh/Documents/Papers3_revised/Library.papers3/Articles/2014/Iandola/arXiv%202014%20Iandola.pdf","file":"arXiv 2014 Iandola.pdf:/Users/yimengzh/Documents/Papers3_revised/Library.papers3/Articles/2014/Iandola/arXiv 2014 Iandola.pdf:application/pdf","uri":"˘rlpapers3://publication/uuid/744F7A4F-AB1A-43F1-8F08-25A025987E62","bibtex":"@article{Iandola:2014tj,\nauthor = {Iandola, F and Moskewicz, M and Karayev, S and Girshick, Ross B and Darrell, Trevor and Keutzer, K},\ntitle = {{DenseNet: Implementing Efficient ConvNet Descriptor Pyramids}},\njournal = {ArXiv e-prints},\nyear = {2014},\nvolume = {cs.CV},\nmonth = apr,\nannote = {essentially, a cheaper method to computing features on region proposals, by computing feature map of whole images at multiple scales, and cropping them to regiona proposals.\n\n\nActually such method dates back to older models such as DPM. Maybe because CNN features are global, not local (like HOG), and thus this method is not implemented before, as it's not accurate (since CNN has pooling, strides, etc;), compared to sliding window feature in HOG.\n\nIt's not elegant, such as aspect ratio, etc. Later on much better methods (Faster RCNN, etc.) are used. People find that a single scale might be sufficient.},\nkeywords = {deep learning},\nread = {Yes},\nrating = {2},\ndate-added = {2017-04-24T15:21:02GMT},\ndate-modified = {2017-04-24T15:32:26GMT},\nurl = {http://arxiv.org/abs/1404.1869},\nlocal-url = {file://localhost/Users/yimengzh/Documents/Papers3_revised/Library.papers3/Articles/2014/Iandola/arXiv%202014%20Iandola.pdf},\nfile = {{arXiv 2014 Iandola.pdf:/Users/yimengzh/Documents/Papers3_revised/Library.papers3/Articles/2014/Iandola/arXiv 2014 Iandola.pdf:application/pdf}},\nuri = {\\url{papers3://publication/uuid/744F7A4F-AB1A-43F1-8F08-25A025987E62}}\n}\n\n\n\n","author_short":["Iandola, F","Moskewicz, M","Karayev, S","Girshick, R. B","Darrell, T.","Keutzer, K"],"key":"Iandola:2014tj","id":"Iandola:2014tj","bibbaseid":"iandola-moskewicz-karayev-girshick-darrell-keutzer-densenetimplementingefficientconvnetdescriptorpyramids-2014","role":"author","urls":{"Paper":"http://arxiv.org/abs/1404.1869"},"keyword":["deep learning"],"metadata":{"authorlinks":{}},"downloads":0,"html":""},"search_terms":["densenet","implementing","efficient","convnet","descriptor","pyramids","iandola","moskewicz","karayev","girshick","darrell","keutzer"],"keywords":["deep learning"],"authorIDs":[],"dataSources":["bzxc3uBcwMv3h47xE","am4cuScX42bDPukDn","KZFesWbmGy4yc4ZLC"]}