My Results on pascal_voc2012 with pretrained FPN_iter_370000.ckpt: CUDA_VISIBLE_DEVICES=0 python ./faster_rcnn/test_net.py --gpu 0 --weights output/FPN_end2end/voc_0712_trainval/FPN_iter_370000.ckpt --imdb voc_2012_test --cfg ./experiments/cfgs/FPN_end2end.yml --network FPN_test

AP for aeroplane = 0.9307 AP for bicycle = 0.2078 AP for bird = 0.7622 AP for boat = 0.6472 AP for bottle = 0.3007 AP for bus = 0.8024 AP for car = 0.1888 AP for cat = 0.9164 AP for chair = 0.1178 AP for cow = 0.9082 AP for diningtable = 0.2955 AP for dog = 0.7173 AP for horse = 0.1252 AP for motorbike = 0.2300 AP for person = 0.6727 AP for pottedplant = 0.3492 AP for sheep = 0.8675 AP for sofa = 0.5945 AP for train = 0.9128 AP for tvmonitor = 0.2478 Mean AP = 0.5397

compared to declared mAP=0.7832 over pascal_voc0712.

alt_opt training:

nohup ./experiments/scripts/FPN_alt_opt.sh 0 FPN_alt_opt pascal_voc0712 --set RNG_SEED 42 TRAIN.SCALES "[800]" > FPN_alt_opt.log 2>&1 &

end2end training:

nohup ./experiments/scripts/FPN_end2end.sh 1 FPN pascal_voc0712 --set RNG_SEED 42 TRAIN.SCALES "[800]" > FPN.log 2>&1 &

tail -f FPN.log

TODO:

1. imporve end2end training result
2. check roi_pooling used interpolation or not
3. fix bugs in alt_opt training