new backbone
parul12111 opened this issue · comments
Hi! Thank you for this amazing repository. I want to build my own backbone and integrate it into rpt, what are the necessary steps and can you please list all the files and directories where I will need to make changes. Thank you!
Hello， I prepare to sort out the whole code and open source , but recently my company is busy, so there is not plenty of time to make training code open source, so there is no training code yet.
But the training code is easy, it is modified from reppoints and siamrpn++
So the first thing you need to do is writing the training code.
If you want to change backbone , you should add the backbone file into reppoints->models->backbone, modify the init.py in reppoints->models->backbone, and also you should modify the config.yaml, according to the output size of feature map, yor should modify '__C.TRAIN.OUTPUT_SIZE' in core->config.py
ok thank you
I am trying to changed the backbone from resnet50 to resnet101.
but when I run eval.py I got this error-
python eval1.py --tracker_path ./results --dataset VOT2018 --tracker_prefix siamreppoints
loading VOT2018: 100%|██████████████████████████████████| 60/60 [00:00<00:00, 78.59it/s, zebrafish1]
eval ar: 0%| | 0/1 [00:00<?, ?it/s]C:\Users\jinda\OneDrive\Desktop\RPT-master\toolkit\utils\statistics.py:54: RuntimeWarning: Mean of empty slice
acc = np.nanmean(overlaps)
C:\Users\jinda\OneDrive\Desktop\RPT-master\toolkit\utils\statistics.py:54: RuntimeWarning: Mean of empty slice
acc = np.nanmean(overlaps)
can u please tell me where i am wrong..
it is not evaluating accuracy and eao
eval ar: 100%|████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 2.97it/s]
eval eao: 0%| | 0/1 [00:00<?, ?it/s]C:\Users\jinda\OneDrive\Desktop\RPT-master\toolkit\utils\statistics.py:54: RuntimeWarning: Mean of empty slice
acc = np.nanmean(overlaps)
eval eao: 100%|███████████████████████████████████████████████████████| 1/1 [00:12<00:00, 12.00s/it]
| Tracker Name | Accuracy | Robustness | Lost Number | EAO |
C:\Users\jinda\OneDrive\Desktop\RPT-master\toolkit\evaluation\ar_benchmark.py:73: RuntimeWarning: Mean of empty slice
accuracy = np.nanmean(overlaps)
| siamreppoints | nan | 16.735 | 3574.0 | 0.000 |
Hello, the backbone resnet50 we used is from siamrpn++, they remove the padding in the 'self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=0, bias=False)', also they modified the stride from 16, 32 of resnet50's layer4 and layer5 to 8, 8. It is described in Chapter3.2 in paper SiamRPN++(https://arxiv.org/abs/1812.11703)