Contents -- Precomputed R-CNN Models

Downloading

• The data package can be downloaded by running the fetch_models.sh script in this directory.

R-CNN detectors trained on PASCAL VOC

• ./rcnn_models/voc_2007/rcnn_model_finetuned
  • The CNN for this detector was fine-tuned on PASCAL VOC 2007 trainval
  • Uses: ./caffe_nets/finetune_voc_2007_trainval_iter_70k
  • This detector was trained on PASCAL VOC 2007 trainval
• ./rcnn_models/voc_2012/rcnn_model_finetuned
  • The CNN for this detector was fine-tuned on PASCAL VOC 2012 train
  • Uses: ./caffe_nets/voc_2012/finetune_voc_2012_train_iter_70k
  • This detector was trained on PASCAL VOC 2012 train (val was not used)
• ./rcnn_models/voc_2007/bbox_regressor_final.mat
  • Bounding box regressor that goes with ./rcnn_models/voc_2007/rcnn_model_finetuned
• ./rcnn_models/voc_2012/bbox_regressor_final.mat
  • Bounding box regressor that goes with ./rcnn_models/voc_2012/rcnn_model_finetuned

Caffe CNNs

• ./caffe_nets/ilsvrc_2012_train_iter_310k
  • Reference network trained for 310k iterations (~66 epochs) on ILSVRC 2012 train
• ./caffe_nets/finetune_voc_2007_trainval_iter_70k
  • Fine-tuned network (initialized from ilsvrc_2012_train_iter_310k) trained for 70k iterations on PASCAL VOC 2007 trainval
• ./caffe_nets/finetune_voc_2012_train_iter_70k
  • Fine-tuned network (initialized from ilsvrc_2012_train_iter_310k) that was trained for 70k iterations on PASCAL VOC 2012 train

CNN fine-tuning notes

• crop mode: warp
• crop padding: 16
• foreground sampling bias: 0.25
• foreground overlap threshold: 0.5
• background overlap threshold: 0.5
• batch size: 128
• base learning rate: 0.001
• window files: see below

Window files used for fine-tuning

• ./window_files/*
• Not included; these can be generated with rcnn_make_window_file.m.

Pre-computed Selective Search boxes

(Downloaded using data/fetch_selective_search_data.sh)

• ./selective_search_boxes/voc_2007_train.mat
• ./selective_search_boxes/voc_2007_val.mat
• ./selective_search_boxes/voc_2007_trainval.mat
• ./selective_search_boxes/voc_2007_test.mat
• ./selective_search_boxes/voc_2012_train.mat
• ./selective_search_boxes/voc_2012_val.mat
• ./selective_search_boxes/voc_2012_trainval.mat
• ./selective_search_boxes/voc_2012_test.mat

Selective Search notes

• The IJCV code was used
• For PASCAL VOC 2007, the pre-computed boxes from the website above were used
• For PASCAL VOC 2010-2012, we computed our own boxes
• "fast mode", which results in about 2k boxes per image, was used