Resnet-101 with fasterRCNN using test_frcnn is not working.

[ignvinay]

Hi , I am trying to get FasterRCNN with resnet-101 running using caffe's cpp code. I specifically use test_frcnn.exe for running this. My command line exeuction looks as below : test_frcnn-d --model C:\Projects\caffe-faster-rcnn\models\FRCNN\res101\test.proto --weights C:\Projects\caffe-faster-rcnn-cpp-windows\examples\FRCNN\resnet-101_rpn_stage1_iter_320000.caffemodel --default_c C:\Projects\caffe-faster-rcnn-cpp-windows\examples\FRCNN\config\voc_config.json --image_root C:\Projects\VOC2007\JPEGImages\ --image_list C:\Projects\caffe-faster-rcnn-cpp-windows\examples\FRCNN\dataset\voc2007.test --out_file C:\Projects\caffe-faster-rcnn-cpp-windows\scripts\build\examples\FRCNN\results\voc2007_test_res101.frcnn --max_per_image 100

  I have taken FasterRCNN with resnet-101 model from link: 

https://onedrive.live.com/?authkey=%21AM0tZPqBFtMejZ8&cid=D75441070ACA1109&id=D75441070ACA1109%211281&parId=root&action=locate

The output of test_frcnn looks like this

I1031 11:28:40.435735 14092 frcnn_api.cpp:47] SET MODEL DONE, ROI POOLING LAYER : roi_pool4 I1031 11:28:40.435735 14092 frcnn_param.cpp:151] == Train Parameters == I1031 11:28:40.436702 14092 frcnn_param.cpp:152] scale : 600.00 I1031 11:28:40.438741 14092 frcnn_param.cpp:153] max_size : 1000 I1031 11:28:40.438741 14092 frcnn_param.cpp:154] batch_size : 128 I1031 11:28:40.439693 14092 frcnn_param.cpp:156] fg_fraction : 0.25 I1031 11:28:40.440692 14092 frcnn_param.cpp:157] fg_thresh : 0.5 I1031 11:28:40.440692 14092 frcnn_param.cpp:158] bg_thresh_hi : 0.5 I1031 11:28:40.440692 14092 frcnn_param.cpp:159] bg_thresh_lo : 0 I1031 11:28:40.441689 14092 frcnn_param.cpp:160] use_flipped : yes I1031 11:28:40.441689 14092 frcnn_param.cpp:162] use_bbox_reg : yes I1031 11:28:40.442687 14092 frcnn_param.cpp:163] bbox_thresh : 0.5 I1031 11:28:40.442687 14092 frcnn_param.cpp:164] snapshot_infix : I1031 11:28:40.442687 14092 frcnn_param.cpp:165] normalize_targets : yes I1031 11:28:40.443683 14092 frcnn_param.cpp:167] rpn_pos_overlap : 0.7 I1031 11:28:40.443683 14092 frcnn_param.cpp:168] rpn_neg_overlap : 0.3 I1031 11:28:40.443683 14092 frcnn_param.cpp:169] clobber_positives : no I1031 11:28:40.444680 14092 frcnn_param.cpp:170] rpn_fg_fraction : 0.5 I1031 11:28:40.444680 14092 frcnn_param.cpp:171] rpn_batchsize : 256 I1031 11:28:40.445677 14092 frcnn_param.cpp:172] rpn_nms_thresh : 0.7 I1031 11:28:40.447674 14092 frcnn_param.cpp:173] rpn_pre_nms_top_n : 12000 I1031 11:28:40.448670 14092 frcnn_param.cpp:174] rpn_post_nms_top_n: 2000 I1031 11:28:40.448670 14092 frcnn_param.cpp:175] rpn_min_size : 16 I1031 11:28:40.448670 14092 frcnn_param.cpp:176] rpn_bbox_inside_weights :1.00, 1.00 I1031 11:28:40.449669 14092 frcnn_param.cpp:177] rpn_positive_weight :-1 I1031 11:28:40.449669 14092 frcnn_param.cpp:178] rpn_allowed_border :0 I1031 11:28:40.450664 14092 frcnn_param.cpp:180] == Test Parameters == I1031 11:28:40.450664 14092 frcnn_param.cpp:181] test_scales : 600.00 I1031 11:28:40.450664 14092 frcnn_param.cpp:182] test_max_size : 1000 I1031 11:28:40.451660 14092 frcnn_param.cpp:183] test_nms : 0.3 I1031 11:28:40.451660 14092 frcnn_param.cpp:184] test_bbox_reg : yes I1031 11:28:40.451660 14092 frcnn_param.cpp:185] test_rpn_nms_thresh : 0.7 I1031 11:28:40.452657 14092 frcnn_param.cpp:186] rpn_pre_nms_top_n : 12000 I1031 11:28:40.452657 14092 frcnn_param.cpp:187] rpn_post_nms_top_n : 300 I1031 11:28:40.453656 14092 frcnn_param.cpp:188] test_rpn_min_sizen : 16 I1031 11:28:40.453656 14092 frcnn_param.cpp:190] == Global Parameters == I1031 11:28:40.453656 14092 frcnn_param.cpp:191] pixel_means[BGR] : 102.98 , 115.947 , 122.772 I1031 11:28:40.454653 14092 frcnn_param.cpp:192] rng_seed : 3 I1031 11:28:40.454653 14092 frcnn_param.cpp:193] eps : 1e-14 I1031 11:28:40.454653 14092 frcnn_param.cpp:194] inf : 1e+08 I1031 11:28:40.456647 14092 frcnn_param.cpp:195] feat_stride : 16 I1031 11:28:40.457644 14092 frcnn_param.cpp:196] anchors_size : 36 I1031 11:28:40.457644 14092 frcnn_param.cpp:197] test_score_thresh : 0.05 I1031 11:28:40.458643 14092 frcnn_param.cpp:198] n_classes : 21 I1031 11:28:40.458643 14092 frcnn_param.cpp:199] iter_test : -1 I1031 11:28:40.459645 14092 test_frcnn.cpp:78] image list : C:\Projects\caffe-faster-rcnn-cpp-windows\examples\FRCNN\dataset\voc2007.test I1031 11:28:40.459645 14092 test_frcnn.cpp:79] output file : C:\Projects\caffe-faster-rcnn-cpp-windows\scripts\build\examples\FRCNN\results\voc2007_test_res101.frcnn I1031 11:28:40.459645 14092 test_frcnn.cpp:80] image root : C:\Projects\VOC2007\JPEGImages
I1031 11:28:40.460638 14092 test_frcnn.cpp:81] max_per_image : 100 I1031 11:30:42.327216 14092 frcnn_api.cpp:81] height: 500 width: 353 E1031 11:30:43.508926 14092 frcnn_api.cpp:98] im_info : 850, 600, 1.69972 E1031 11:30:43.560787 14092 frcnn_api.cpp:8] img_in (CHW) : 3, 850, 600 E1031 11:30:52.687629 14092 frcnn_api.cpp:52] FORWARD BEGIN E1031 11:31:22.743995 14092 frcnn_proposal_layer.cpp:47] ========== enter proposal layer E1031 11:31:34.242405 14092 frcnn_proposal_layer.cpp:89] ========== generate anchors E1031 11:31:34.382033 14092 frcnn_proposal_layer.cpp:126] ========== after clip and remove size < threshold box 18120 E1031 11:31:34.495728 14092 frcnn_proposal_layer.cpp:135] ========== apply nms, pre nms number is : 12000 E1031 11:31:36.456616 14092 frcnn_proposal_layer.cpp:153] rpn number after nms: 300 E1031 11:31:36.458629 14092 frcnn_proposal_layer.cpp:155] ========== copy to top E1031 11:31:36.533031 14092 frcnn_proposal_layer.cpp:174] ========== exit proposal layer E1031 11:32:13.613378 14092 frcnn_api.cpp:59] FORWARD END, Loss : 0 I1031 11:32:16.731925 14092 test_frcnn.cpp:121] Handle 1 th image : 000001.jpg, with image_thresh : 0, 0 -> 0 boxes I1031 11:32:17.777040 14092 frcnn_api.cpp:81] height: 500 width: 335 E1031 11:32:18.557649 14092 frcnn_api.cpp:98] im_info : 896, 600, 1.79104 E1031 11:32:18.627769 14092 frcnn_api.cpp:8] img_in (CHW) : 3, 896, 600 E1031 11:32:19.490888 14092 frcnn_api.cpp:52] FORWARD BEGIN

Its not able to detect the objects in image. Whereas if i feed fasterRCNN with VGG16, it shows objects. Am i doing anything wrong here ?, Please help.

[ignvinay]

Any ideas , please !