ralphmao / patchnet

PyTorch-PatchNet

Overview

This repo contains PatchNet model description in PyTorch and its training/evaluation with SiameseFC. It heavily relies on siamfc-pytorch.

• Demo
• Prerequisite
• How to use

Demo

From left to right: Skip frame - interleave siamfc with patchnet for speedup. Boosted - combine siamfc's prediction scores and patchnet's local similarity scores for more robust tracking. SiamFC - the base line.

All three methods are tested with PyTorch-1.4 on Nvidia Jetson Nano.

Prerequisite

• Python 3
• PyTorch >= 1.2
• GOT-10k dataset (for training only)

How to use

Both trained models of SiamFC and PatchNet have already been stored under ./pretrained/

Demo

python3 scripts/demo_skipframe.py

Interleave SiamFC and PatchNet on OTB2015 dataset

python3 scripts/test_<name>.py

Interleave RFCN ResNet-101 and PatchNet

python3 scripts/vod_track.py data/VID/ILSVRC2015_val_00142000 data/VID/rfcn_res101_val_00142000.txt --visualize --conf 0.5

Train PatchNet model

python3 scripts/train_patchnet.py --dataset <got_10k_root> --save pretrained/<my_fancy_model>

patchnet's Issues

Low fps on jetson nano

Hi. I am trying to run PatchNet on Jetson Nano. Measured fps is aroud 15-16(demo_skipframe.py) when number of skipframes is set to 1000(so only PatchNet is running). Gpu load is <=50%. What can be done?
Jetpack-4.6.1
CUDA-10.2
Python-3.6
Torch-1.10

fourier_feature.py

hi @RalphMao ..
fourier_feature is not wotk with pytorch 1.7.0 because the function torch.rfft is removed

can you help my to edit your code to work with pytorch 1.7.0

thanks.....
@Sibozhu .....