search

MhLiao / TextBoxes_plusplus

TextBoxes++: A Single-Shot Oriented Scene Text Detector
Other
953 stars 279 forks source link

Can not get a good performance on dataset RCTW-17. #45

Closed justttry closed 5 years ago

justttry commented 6 years ago

I trained the model on dataset RCTW-17 for 5 days, but the det_eval is still about 0.6. The loss could not decrease any more. Any suggestion?

below is the print. I print the det_eval per 1000 iters.

I0404 17:52:12.688493 1306 solver.cpp:543] Test net output #0: detection_eval = 0.58883 I0404 17:52:17.242920 1306 solver.cpp:243] Iteration 34000, loss = 2.45406 I0404 17:52:17.242971 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.45406 (* 1 = 2.45406 loss) I0404 17:52:17.683764 1306 sgd_solver.cpp:138] Iteration 34000, lr = 0.0001

I0404 21:02:53.277609 1306 solver.cpp:433] Iteration 35000, Testing net (#0) I0404 21:02:53.277750 1306 net.cpp:693] Ignoring source layer mbox_loss I0404 21:03:32.067988 1306 solver.cpp:543] Test net output #0: detection_eval = 0.635565 I0404 21:03:34.451148 1306 solver.cpp:243] Iteration 35000, loss = 2.24144 I0404 21:03:34.451225 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.9662 (* 1 = 1.9662 loss) I0404 21:03:35.548418 1306 sgd_solver.cpp:138] Iteration 35000, lr = 0.0001

I0405 00:46:20.712731 1306 solver.cpp:433] Iteration 36000, Testing net (#0) I0405 00:46:20.712879 1306 net.cpp:693] Ignoring source layer mbox_loss I0405 00:46:37.906997 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0405 00:46:59.443310 1306 solver.cpp:543] Test net output #0: detection_eval = 0.611436 I0405 00:47:03.187078 1306 solver.cpp:243] Iteration 36000, loss = 2.218 I0405 00:47:03.187140 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.50099 (* 1 = 2.50099 loss) I0405 00:47:03.717782 1306 sgd_solver.cpp:138] Iteration 36000, lr = 0.0001

I0405 04:28:44.928535 1306 solver.cpp:433] Iteration 37000, Testing net (#0) I0405 04:28:44.928661 1306 net.cpp:693] Ignoring source layer mbox_loss I0405 04:29:24.820894 1306 solver.cpp:543] Test net output #0: detection_eval = 0.604593 I0405 04:29:28.307677 1306 solver.cpp:243] Iteration 37000, loss = 2.17795 I0405 04:29:28.307729 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.18753 (* 1 = 2.18753 loss) I0405 04:29:28.307762 1306 sgd_solver.cpp:138] Iteration 37000, lr = 0.0001

I0405 08:09:37.835278 1306 solver.cpp:433] Iteration 38000, Testing net (#0) I0405 08:09:37.835407 1306 net.cpp:693] Ignoring source layer mbox_loss I0405 08:10:10.449904 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0405 08:10:14.990397 1306 solver.cpp:543] Test net output #0: detection_eval = 0.627953 I0405 08:10:18.733402 1306 solver.cpp:243] Iteration 38000, loss = 2.21655 I0405 08:10:18.733454 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.35411 (* 1 = 2.35411 loss) I0405 08:10:18.733497 1306 sgd_solver.cpp:138] Iteration 38000, lr = 0.0001

I0405 13:15:23.770017 1306 solver.cpp:433] Iteration 39000, Testing net (#0) I0405 13:15:23.770084 1306 net.cpp:693] Ignoring source layer mbox_loss I0405 13:16:04.455183 1306 solver.cpp:543] Test net output #0: detection_eval = 0.594352 I0405 13:16:07.660732 1306 solver.cpp:243] Iteration 39000, loss = 2.18507 I0405 13:16:07.660820 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.15974 (* 1 = 2.15974 loss) I0405 13:16:07.660889 1306 sgd_solver.cpp:138] Iteration 39000, lr = 0.0001

I0405 17:02:08.661116 1306 solver.cpp:433] Iteration 40000, Testing net (#0) I0405 17:02:08.661207 1306 net.cpp:693] Ignoring source layer mbox_loss I0405 17:02:49.289695 1306 solver.cpp:543] Test net output #0: detection_eval = 0.588826 I0405 17:02:51.784826 1306 solver.cpp:243] Iteration 40000, loss = 2.19788 I0405 17:02:51.784868 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.65966 (* 1 = 2.65966 loss) I0405 17:02:52.473186 1306 sgd_solver.cpp:47] MultiStep Status: Iteration 40000, step = 1 I0405 17:02:52.473227 1306 sgd_solver.cpp:138] Iteration 40000, lr = 1e-05

I0405 21:15:27.636586 1306 solver.cpp:433] Iteration 41000, Testing net (#0) I0405 21:15:27.636674 1306 net.cpp:693] Ignoring source layer mbox_loss I0405 21:15:33.403995 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0405 21:16:08.100375 1306 solver.cpp:543] Test net output #0: detection_eval = 0.607239 I0405 21:16:10.347666 1306 solver.cpp:243] Iteration 41000, loss = 2.05813 I0405 21:16:10.347721 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.09432 (* 1 = 2.09432 loss) I0405 21:16:11.122637 1306 sgd_solver.cpp:138] Iteration 41000, lr = 1e-05

I0406 01:03:01.890405 1306 solver.cpp:433] Iteration 42000, Testing net (#0) I0406 01:03:01.890519 1306 net.cpp:693] Ignoring source layer mbox_loss I0406 01:03:40.885931 1306 solver.cpp:543] Test net output #0: detection_eval = 0.624692 I0406 01:03:44.217214 1306 solver.cpp:243] Iteration 42000, loss = 2.01631 I0406 01:03:44.217658 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.71339 (* 1 = 2.71339 loss) I0406 01:03:44.217883 1306 sgd_solver.cpp:138] Iteration 42000, lr = 1e-05

I0406 05:10:18.305634 1306 solver.cpp:433] Iteration 43000, Testing net (#0) I0406 05:10:18.305721 1306 net.cpp:693] Ignoring source layer mbox_loss I0406 05:10:33.529862 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0406 05:10:57.730491 1306 solver.cpp:543] Test net output #0: detection_eval = 0.597785 I0406 05:11:00.188812 1306 solver.cpp:243] Iteration 43000, loss = 1.9701 I0406 05:11:00.188872 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.00485 (* 1 = 2.00485 loss) I0406 05:11:00.853934 1306 sgd_solver.cpp:138] Iteration 43000, lr = 1e-05

I0406 08:45:31.997023 1306 solver.cpp:433] Iteration 44000, Testing net (#0) I0406 08:45:31.997119 1306 net.cpp:693] Ignoring source layer mbox_loss I0406 08:46:11.371773 1306 solver.cpp:543] Test net output #0: detection_eval = 0.60564 I0406 08:46:14.174206 1306 solver.cpp:243] Iteration 44000, loss = 2.0197 I0406 08:46:14.174255 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.05228 (* 1 = 2.05228 loss) I0406 08:46:14.756858 1306 sgd_solver.cpp:138] Iteration 44000, lr = 1e-05

I0406 12:15:49.126451 1306 solver.cpp:433] Iteration 45000, Testing net (#0) I0406 12:15:49.126693 1306 net.cpp:693] Ignoring source layer mbox_loss I0406 12:16:22.482960 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0406 12:16:27.758494 1306 solver.cpp:543] Test net output #0: detection_eval = 0.584604 I0406 12:16:30.284601 1306 solver.cpp:243] Iteration 45000, loss = 1.96844 I0406 12:16:30.284667 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.69729 (* 1 = 1.69729 loss) I0406 12:16:31.092077 1306 sgd_solver.cpp:138] Iteration 45000, lr = 1e-05

I0406 15:25:25.520157 1306 solver.cpp:433] Iteration 46000, Testing net (#0) I0406 15:25:25.520244 1306 net.cpp:693] Ignoring source layer mbox_loss I0406 15:26:05.611193 1306 solver.cpp:543] Test net output #0: detection_eval = 0.613347 I0406 15:26:08.824692 1306 solver.cpp:243] Iteration 46000, loss = 1.9416 I0406 15:26:08.824733 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.83413 (* 1 = 1.83413 loss) I0406 15:26:08.824764 1306 sgd_solver.cpp:138] Iteration 46000, lr = 1e-05

I0406 18:27:18.800853 1306 solver.cpp:433] Iteration 47000, Testing net (#0) I0406 18:27:18.800943 1306 net.cpp:693] Ignoring source layer mbox_loss I0406 18:27:56.227263 1306 solver.cpp:543] Test net output #0: detection_eval = 0.627901 I0406 18:27:59.330350 1306 solver.cpp:243] Iteration 47000, loss = 1.93108 I0406 18:27:59.330410 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.17465 (* 1 = 2.17465 loss) I0406 18:27:59.330467 1306 sgd_solver.cpp:138] Iteration 47000, lr = 1e-05

I0406 21:27:43.450727 1306 solver.cpp:433] Iteration 48000, Testing net (#0) I0406 21:27:43.450819 1306 net.cpp:693] Ignoring source layer mbox_loss I0406 21:28:22.387928 1306 solver.cpp:543] Test net output #0: detection_eval = 0.652113 I0406 21:28:24.734930 1306 solver.cpp:243] Iteration 48000, loss = 1.95157 I0406 21:28:24.735054 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.1868 (* 1 = 2.1868 loss) I0406 21:28:25.480062 1306 sgd_solver.cpp:138] Iteration 48000, lr = 1e-05

I0407 00:30:41.377358 1306 solver.cpp:433] Iteration 49000, Testing net (#0) I0407 00:30:41.377449 1306 net.cpp:693] Ignoring source layer mbox_loss I0407 00:31:18.830639 1306 solver.cpp:543] Test net output #0: detection_eval = 0.617053 I0407 00:31:21.245787 1306 solver.cpp:243] Iteration 49000, loss = 1.9707 I0407 00:31:21.245849 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.75377 (* 1 = 1.75377 loss) I0407 00:31:21.680156 1306 sgd_solver.cpp:138] Iteration 49000, lr = 1e-05

I0407 03:30:56.405927 1306 solver.cpp:433] Iteration 50000, Testing net (#0) I0407 03:30:56.406105 1306 net.cpp:693] Ignoring source layer mbox_loss I0407 03:31:01.051456 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0407 03:31:34.187222 1306 solver.cpp:543] Test net output #0: detection_eval = 0.596666 I0407 03:31:36.562472 1306 solver.cpp:243] Iteration 50000, loss = 1.93141 I0407 03:31:36.562527 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.87467 (* 1 = 1.87467 loss) I0407 03:31:37.268019 1306 sgd_solver.cpp:138] Iteration 50000, lr = 1e-05

I0407 06:32:31.886972 1306 solver.cpp:433] Iteration 51000, Testing net (#0) I0407 06:32:31.887068 1306 net.cpp:693] Ignoring source layer mbox_loss I0407 06:33:11.281626 1306 solver.cpp:543] Test net output #0: detection_eval = 0.682606 I0407 06:33:13.594921 1306 solver.cpp:243] Iteration 51000, loss = 1.93064 I0407 06:33:13.594975 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.09862 (* 1 = 2.09862 loss) I0407 06:33:14.407634 1306 sgd_solver.cpp:138] Iteration 51000, lr = 1e-05

I0407 09:35:26.773157 1306 solver.cpp:433] Iteration 52000, Testing net (#0) I0407 09:35:26.773277 1306 net.cpp:693] Ignoring source layer mbox_loss I0407 09:35:46.128684 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0407 09:36:05.563726 1306 solver.cpp:543] Test net output #0: detection_eval = 0.621031 I0407 09:36:08.164755 1306 solver.cpp:243] Iteration 52000, loss = 1.95618 I0407 09:36:08.164818 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.84878 (* 1 = 1.84878 loss) I0407 09:36:08.665041 1306 sgd_solver.cpp:138] Iteration 52000, lr = 1e-05

I0407 12:39:53.927330 1306 solver.cpp:433] Iteration 53000, Testing net (#0) I0407 12:39:53.927418 1306 net.cpp:693] Ignoring source layer mbox_loss I0407 12:40:34.113943 1306 solver.cpp:543] Test net output #0: detection_eval = 0.624961 I0407 12:40:36.511904 1306 solver.cpp:243] Iteration 53000, loss = 1.92391 I0407 12:40:36.511981 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.79335 (* 1 = 1.79335 loss) I0407 12:40:36.951359 1306 sgd_solver.cpp:138] Iteration 53000, lr = 1e-05

I0407 15:41:31.221352 1306 solver.cpp:433] Iteration 54000, Testing net (#0) I0407 15:41:31.221436 1306 net.cpp:693] Ignoring source layer mbox_loss I0407 15:42:00.742380 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0407 15:42:09.181475 1306 solver.cpp:543] Test net output #0: detection_eval = 0.647486 I0407 15:42:11.437472 1306 solver.cpp:243] Iteration 54000, loss = 1.90603 I0407 15:42:11.437587 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.02106 (* 1 = 2.02106 loss) I0407 15:42:12.188097 1306 sgd_solver.cpp:138] Iteration 54000, lr = 1e-05

I0407 18:48:10.112860 1306 solver.cpp:433] Iteration 55000, Testing net (#0) I0407 18:48:10.112948 1306 net.cpp:693] Ignoring source layer mbox_loss I0407 18:48:49.613601 1306 solver.cpp:543] Test net output #0: detection_eval = 0.63675 I0407 18:48:52.417958 1306 solver.cpp:243] Iteration 55000, loss = 1.93508 I0407 18:48:52.418064 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.4041 (* 1 = 2.4041 loss) I0407 18:48:53.797850 1306 sgd_solver.cpp:138] Iteration 55000, lr = 1e-05

I0407 21:55:46.550694 1306 solver.cpp:433] Iteration 56000, Testing net (#0) I0407 21:55:46.550786 1306 net.cpp:693] Ignoring source layer mbox_loss I0407 21:56:25.088029 1306 solver.cpp:543] Test net output #0: detection_eval = 0.611102 I0407 21:56:27.633196 1306 solver.cpp:243] Iteration 56000, loss = 1.87605 I0407 21:56:27.633272 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.78036 (* 1 = 1.78036 loss) I0407 21:56:28.232192 1306 sgd_solver.cpp:138] Iteration 56000, lr = 1e-05

I0408 01:05:07.849153 1306 solver.cpp:433] Iteration 57000, Testing net (#0) I0408 01:05:07.849238 1306 net.cpp:693] Ignoring source layer mbox_loss I0408 01:05:08.027675 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0408 01:05:45.337280 1306 solver.cpp:543] Test net output #0: detection_eval = 0.61578 I0408 01:05:47.763682 1306 solver.cpp:243] Iteration 57000, loss = 1.87234 I0408 01:05:47.763732 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.73128 (* 1 = 1.73128 loss) I0408 01:05:48.229641 1306 sgd_solver.cpp:138] Iteration 57000, lr = 1e-05

I0408 04:28:34.337592 1306 solver.cpp:433] Iteration 58000, Testing net (#0) I0408 04:28:34.337704 1306 net.cpp:693] Ignoring source layer mbox_loss I0408 04:29:14.902088 1306 solver.cpp:543] Test net output #0: detection_eval = 0.642386 I0408 04:29:17.686023 1306 solver.cpp:243] Iteration 58000, loss = 1.85526 I0408 04:29:17.686065 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.76768 (* 1 = 1.76768 loss) I0408 04:29:18.140614 1306 sgd_solver.cpp:138] Iteration 58000, lr = 1e-05

I0408 08:14:14.440521 1306 solver.cpp:433] Iteration 59000, Testing net (#0) I0408 08:14:14.440623 1306 net.cpp:693] Ignoring source layer mbox_loss I0408 08:14:26.626003 1306 blocking_queue.cpp:50] Data layer prefetch queue empty I0408 08:14:53.532229 1306 solver.cpp:543] Test net output #0: detection_eval = 0.614253 I0408 08:14:56.856050 1306 solver.cpp:243] Iteration 59000, loss = 1.92192 I0408 08:14:56.856138 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.69259 (* 1 = 1.69259 loss) I0408 08:14:56.856176 1306 sgd_solver.cpp:138] Iteration 59000, lr = 1e-05

I0408 12:02:09.806546 1306 solver.cpp:433] Iteration 60000, Testing net (#0) I0408 12:02:09.806643 1306 net.cpp:693] Ignoring source layer mbox_loss I0408 12:02:57.607538 1306 solver.cpp:543] Test net output #0: detection_eval = 0.611193 I0408 12:03:03.562712 1306 solver.cpp:243] Iteration 60000, loss = 1.91013 I0408 12:03:03.562777 1306 solver.cpp:259] Train net output #0: mbox_loss = 2.27964 ( 1 = 2.27964 loss) I0408 12:03:07.462204 1306 sgd_solver.cpp:138] Iteration 60000, lr = 1e-05 I0408 12:37:51.408920 1306 solver.cpp:243] Iteration 60100, loss = 1.8492 I0408 12:37:51.409126 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.80072 ( 1 = 1.80072 loss) I0408 12:38:11.501283 1306 sgd_solver.cpp:138] Iteration 60100, lr = 1e-05 I0408 13:12:53.471482 1431 blocking_queue.cpp:50] Data layer prefetch queue empty I0408 13:18:36.838145 1306 solver.cpp:243] Iteration 60200, loss = 1.91048 I0408 13:18:36.838371 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.98924 ( 1 = 1.98924 loss) I0408 13:19:00.084318 1306 sgd_solver.cpp:138] Iteration 60200, lr = 1e-05 I0408 13:55:33.791471 1306 solver.cpp:243] Iteration 60300, loss = 1.87662 I0408 13:55:33.791700 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.9425 ( 1 = 1.9425 loss) I0408 13:55:47.231323 1306 sgd_solver.cpp:138] Iteration 60300, lr = 1e-05 I0408 14:35:58.646726 1306 solver.cpp:243] Iteration 60400, loss = 1.87251 I0408 14:35:58.647043 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.55061 ( 1 = 1.55061 loss) I0408 14:36:08.683598 1306 sgd_solver.cpp:138] Iteration 60400, lr = 1e-05 I0408 15:13:43.952873 1306 solver.cpp:243] Iteration 60500, loss = 1.91905 I0408 15:13:43.953128 1306 solver.cpp:259] Train net output #0: mbox_loss = 1.85035 ( 1 = 1.85035 loss) I0408 15:14:05.191656 1306 sgd_solver.cpp:138] Iteration 60500, lr = 1e-05

MhLiao commented 6 years ago

The det_eval is the "mean average precision", which is not indicated as "F-measure". Did you try to test the performance using the official evaluation protocol?

justttry commented 6 years ago

no, I didnot. I just tested the trained model on training dataset. some pictures could be predicted prefectly, but not all the dataset. the good one like this, good and the bad one like this, bad