Using pascal voc 0712 trainval, faster_rcnn_end2end.sh get higher results 77.1 compared to 73.2on 2007 test

[QishuoLu]

Using pascal voc 0712 trainval, faster_rcnn_end2end.sh get higher results 77.1 compared to 73.2 on 2007 test: AP for aeroplane = 0.7876 AP for bicycle = 0.8118 AP for bird = 0.7623 AP for boat = 0.7207 AP for bottle = 0.6293 AP for bus = 0.8834 AP for car = 0.8818 AP for cat = 0.8854 AP for chair = 0.6011 AP for cow = 0.8184 AP for diningtable = 0.7276 AP for dog = 0.8589 AP for horse = 0.8572 AP for motorbike = 0.7838 AP for person = 0.7899 AP for pottedplant = 0.5071 AP for sheep = 0.7778 AP for sofa = 0.7574 AP for train = 0.8372 AP for tvmonitor = 0.7434 Mean AP = 0.7711

[wk910930]

Dear @QishuoLu , your results looks promising. Did you modify the configures during the TRAINING/TESTING? I cannot get such higher results. Could you please share your configures? Thanks!

[QishuoLu]

Dear @wk910930 my solver.ptototxt configures are as follows: train_net: "models/pascal_voc/VGG16/faster_rcnn_end2end/train.prototxt" base_lr: 0.001 lr_policy: "step" gamma: 0.1 stepsize: 80000 display: 20 average_loss: 100

iter_size: 1

momentum: 0.9 weight_decay: 0.0005

We disable standard caffe solver snapshotting and implement our own snapshot

function

snapshot: 0

We still use the snapshot prefix, though

snapshot_prefix: "vgg16_faster_rcnn" iter_size: 2

and the total stepsize is 11000 configured in faster_rcnn_end2end.sh