researchmm / lighttrack Goto Github PK

感谢你非常棒的工作《LightTrack: Finding Lightweight Neural Networks for Object Tracking via One-Shot Architecture Search》，但是网上好像现在找不到这篇论文，可以给个论文的下载链接吗?另外，可以尽快公布代码吗？

Can you tell me about config of model 790M ?

Hi, I'm interested in SOT, especially efficient SOT algorithm. Thanks for your awesome work.
Could you tell me how to run this repo on my own videos?

Hi, i read the paper again, in section 4.3, it confuse me that how can you make sure that the best backbone architecture(path) can be sampled in the pre-training stage and the sampled path is different from the previous paths?
Since there are about 78 billion possible backbones(paths), and i think it's impossible to pre-train them all and maybe the actual best backbone(path) was not been sampled, so what if this situation was really happend during your expirement?
Thanks, looking for your reply.

很不错的工作，支持； 另外想请教一下 在安卓平台部署的代码使用什么呢 可以分享学习实测一下吗？

how to set the value of threshold? when score lower than the threshold is lost targrt

Hello ,your work is very wonderful and testing in my video achieved a good performance. One frame inference time is about 13ms and ocean is about 15ms in 2080Ti. How to raise the fps? Thanks for your open-source codes and methods and look forward your prompt reply.

Is it MIT license like the STARK tracker from your group?
Great solution BTW :)

downloaded from https://www.votchallenge.net/vot2019/dataset.html

VOT2019 dataset preview

but did not have list.txt

FileNotFoundError: [Errno 2] No such file or directory: 'LightTrack/lib/utils/../../dataset/VOT2019/list.txt'

the confidence of the bbox,when I run some other datasets,I find some FP.

Is there a pretrained weight trained on ImageNet for searched subnet ? I would like to use it to start my track-model's training on the tracking dataset. thanks.

https://github.com/wuhongsheng/ncnn-android-lighttrack

作者你好，请问能否加你微信一下呢？有些问题想咨询一下，我的邮箱是[email protected]。比较急切，还望作者看到之后能回我一下，谢谢

Thanks for your work. I tried to repeat the result in the video testing part (1 video) to get the finished video with the box, but something went wrong with me. I used google collab as my medium. I would like a code that loads the model and then passes the video through itself and returns the box.

https://colab.research.google.com/drive/1JpXUueKsrjOb6LJCy1v-GhcPEDhE8SPG?usp=sharing

when i add --video /path/ to test a video faced error : KeyError: '/dev/video0'
anyone know about it?

Hi! Thank you for publishing your code.

In your paper, you have mentioned various training-related choices/hyperparameters (e.g., learning rate, number of epochs). Your tracker evaluation script has parameters related to the post-processing of bounding-box output by the tracker (penalty_k, window_influence, lr).

Can you please suggest how did you arrive at these hyperparameter values? Was it fine-tuned using the test-set itself?

After testing your work on UAV123 dataset and evaluating the result with OPE benchmark the result is have very low accuracy
Is there any wrong in your code or paper
.....

Hi, it's a very interesting and excellent work, and I am a beginner in NAS, I'm very interesting in the pre-training of backbone supernet, but I didn't find this part in your released code, did you not release this part or it's just my fault?
Thanks, looking for your reply.

Rockchip npu inference version
https://github.com/researchmm/LightTrack/issues

Can I use the provided pre-trained model to track objects in a video recorded by myself? Or do I need to retrain the model based on the content of the recorded video?

agility
ball2
basketball
bolt1
ants1
ball3
birds1
...

content above not ok to run

Hello dear.
I want to use your code for train my own dataset but you have not put training code in repository. would you please put training code?

I noticed that the method of cross-correlation is different from the general siam tracker. In the code, the matrix multiplication method is used as the pixel cross-correlation, and the channel separable convolution method is generally used as the channel cross-correlation. What are the advantages of using pixel correlation?
Looking forward to your reply, thank you very much!

The paper is similar to siamfc++, except that the network layer is changed. Why is it so robust that it is better than online updates, such as dimp. How does this offline tracking tracker distinguish distractor?

Does your open source code support Windows system?

https://github.com/Z-Xiong/LightTrack-ncnn

hello, it's a wonderful work and I wonder when will release the training code?

Hello, your work is shocking, please ask how the code is trained, can you improve it in the readme?

Hi
Great great work! I also ran on other videos and the results are surprisingly well for such a fast network
How did you convert the network to DSP/GPU? is it a single network or split into a few sections?
Where is the correlation performed?
Does the runtime include pre/post processing?
Also, can you please share the file you ran?
Thanks!

你好，我这边把模型部署在了Jetson agx xavier上了，但是测试了在VOT2019上的FPS很低，只有10 FPS左右，还行请问一下这是怎么回事呢？

Hello, how do I run tracker on Snapdragon 845 GPU? Can you share this part of the code? Or describe the technical roadmap in detail.

Hi,

Thank you for your great work.
Could you please share your official raw results on tracking datasets and benchmarks which are used in your paper?

what exactly is this path?
path_name=back_04502514044521042540+cls_211000022+reg_100000111_ops_32

On running: bash tracking/reproduce_vot2019.sh

===> init Siamese <====
load pretrained model from snapshot/LightTrackM/LightTrackM.pth
Traceback (most recent call last):
File "tracking/test_lighttrack.py", line 181, in
main()
File "tracking/test_lighttrack.py", line 165, in main
siam_net = load_pretrain(siam_net, args.resume)
File "/home/nishat/adasi/tracking/LightTrack/tracking/../lib/utils/utils.py", line 702, in load_pretrain
ckpt = torch.load(ckpt_path, map_location=lambda storage, loc: storage(device))
NameError: name 'ckpt_path' is not defined
VOT2019
loading VOT2019: 100%|██████████████████████████████████| 60/60 [00:00<00:00, 61.31it/s, zebrafish1]
eval ar: 100%|████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 15.58it/s]
eval eao: 0%| | 0/1 [00:00<?, ?it/s]
multiprocessing.pool.RemoteTraceback:
"""
Traceback (most recent call last):
File "/home/nishat/miniconda3/envs/lighttrack/lib/python3.6/multiprocessing/pool.py", line 119, in worker
result = (True, func(*args, **kwds))
File "/home/nishat/adasi/tracking/LightTrack/lib/eval_toolkit/bin/../../eval_toolkit/pysot/evaluation/eao_benchmark.py", line 47, in eval
eao = self._calculate_eao(tracker_name, self.tags)
File "/home/nishat/adasi/tracking/LightTrack/lib/eval_toolkit/bin/../../eval_toolkit/pysot/evaluation/eao_benchmark.py", line 110, in _calculate_eao
max_len = max([len(x) for x in all_overlaps])
ValueError: max() arg is an empty sequence
"""

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
File "lib/eval_toolkit/bin/eval.py", line 148, in
trackers), desc='eval eao', total=len(trackers), ncols=100):
File "/home/nishat/miniconda3/envs/lighttrack/lib/python3.6/site-packages/tqdm/std.py", line 1178, in iter
for obj in iterable:
File "/home/nishat/miniconda3/envs/lighttrack/lib/python3.6/multiprocessing/pool.py", line 735, in next
raise value
ValueError: max() arg is an empty sequence

你好，请问一下代码中install.sh里cd lib/models/EfficientNet-PyTorch这句中的EfficientNet-PyTorch怎么没有啊？是需要自己去git clone下来吗？如果需要自己的git clone下来的话，是下载哪个版本的EfficientNet-PyTorch，还麻烦给个链接呢，非常感谢。