judasdie / sots Goto Github PK

Great job!!! Thank you very much!!!
And when will u release the new version of CSTrack?

When will SiamMOT's code be released?

Many thanks for your wonderful work and the code released. But I did not know clearly how the installation is to run CSTrack since only the installation of python or pytorch like are statemented, there seems no description about 'python setup.py install' or 'python install .' like operation for CSTrack. The problem I met, namly 'No module named 'cython_bbox' ', seems due to I did not run 'python setup.py' like, but I did not know which one is needed for CSTrack.

how to export a trained model from PyTorch to ONNX and TorchScript formats.

It is very kind of u if u release a .py for video demo vision，come on!

您好！
我在训练CSTtrack时，按照教程放置了应有的所有训练集，并在json文件中修改了路径。但运行train_cstrack.py时报错：Exception: Error loading data from ../lib/dataset/mot/data/mot15.val: G:/datasets\MOT15/labels_with_ids/train/ADL-Rundle-6/img1/000001.txt not found.
请问我该从哪里得到MOT15/labels_with_ids文件呢？另外，明明没有在MOT15上训练，为什么需要这个数据集呢？
感谢！

你好，我想请教下如何在cstrack中使用自己的检测结果呢，我注意到您在test_cstrack_panda.py中给出了相应参数以便于融合自己的det结果，但是因为panda数据集本身图片分辨率很大，您加入了分割图片的机制，我想用于mot20这样的数据集做一些测试，因为其本身图片就1920*1080，不想延用分割，那么我大概需要修改哪些文件呢，或者能否在test_cstrack.py加入一些代码来融合自己的det结果呢？🐱‍🚀

请问automatch代码会开源吗

I run the command "python train_cstrack.py", and weights dir has yolov5l.pt file. The problem is occured.

Hello, I have a question. Can CSTrack separate detection and reid? For example, can my detection part be changed to yolov4, yolov3? Thanks for the answer!

Hi, do you opensource the re check module of CSTrackV2?

How can I use my own video as a demo for CSTrack testing?

作者您好，我使用您提供的预训练权重cstrack.pt验证在MOT15训练集上的性能，参数nms_thres=0.6，conf_thres=0.5，img_size = (1088, 608)，评估code使用https://github.com/JonathonLuiten/TrackEval，但是获得的效果很差MOTA只有54%，您能告诉我可能是什么原因吗？

你好，我按照readme处理了got10k的数据集，依然找不到all.json这个文件，代码中也没找到涉及all.json的位置。

Hi, I am interested in trying CSTrack/OMC for my use case. I am using triton server for inference. Does CSTrack/OMC support onnx or trt export? Just wanted to confirm with you beforehand. Thanks!

你好，我用你提供的cstrack_panda.pt, 来跑ICCV 2021 Workshop GigaVision I: Multi-Object Tracking这个比赛的测试集，把跑出来的结果框画回到原图很多是正确的， 但是为什么提交文件没分数，有时候有分数了显示0分？ 这个问题可能不是模型、代码相关的问题，因为我看见你也参加了这个比赛，所以想问问你，

使用GOT10K测试是，experiments/test/ 路径下缺少文件 GOT10K/Ocean.yaml

在测试 load dataset 过程中，

你好，@JudasDie

关于AutoMatch，我有几个问题想请教一下。

问题一：
根据你的代码，在L2Mregression和L2Mclassification的初始化中，都已经确定了使用哪两个算子。比如在L2Mregression的__init__函数中，设置了self.LTM = LTM(inchannel=256, used=[5, 6])。请问此处的[5, 6]是根据NAS（论文中提到的第一个训练阶段？）得到的嘛？

问题二：
在如问题一提到的__init__函数下，是不是意味着，你提供的训练指令“python tracking/onekey.py --cfg experiments/AutoMatch.yaml”只包含了第二个训练阶段（因为使用哪两个匹配算子已经固定了）？

问题三：
如果我想要进行第一个阶段的训练，该怎么实现呢？

非常感谢！

Now, I keep trying to train a CSTrack model.
I set the batch_size=2,4,6,8 and input_size=640 but, I can't train it because of CUDA out of memory error.
only a setting(batch_size=1 and input_size=640) operated well...
(mainly, CUDA out of error, my computer is set to 2060 super * 2, and used to train on only 1 GPU)

I don't know exactly what is wrong..

I think the CSTrack model is trained on 2080TI with batch_size=8 or 10 and input_size=640.

is it right..?!

Scaled weight_decay = 0.0004921875
Optimizer groups: 135 .bias, 148 conv.weight, 121 other
Traceback (most recent call last):
File "train_omc.py", line 742, in
main(opt)
File "train_omc.py", line 640, in main
train(opt.hyp, opt, device)
File "train_omc.py", line 266, in train
optimizer.load_state_dict(ckpt['optimizer'])
File "/root/miniconda3/envs/omc/lib/python3.8/site-packages/torch/optim/optimizer.py", line 146, in load_state_dict
raise ValueError("loaded state dict contains a parameter group "
ValueError: loaded state dict contains a parameter group that doesn't match the size of optimizer's group

训练使用单卡2080ti，batch 6

如果跳过optimizer的话，载入siammot也会报错

您好，我用crowdhuman来训练cstrack模型的时候，训练速度很慢，gpu使用率波动很大，一下百分之百，一下百分之零，我尝试调整num_work数量，但是cpu使用率始终只用了一个核，不知道要怎么解决这个问题；第二个问题是我用yolov5l做预训练模型，但训练效果很不理想，recall一直降低，不知道要怎么解决这个问题；第三个问题是您网站上提供的yolov5l模型和cstrack.pt模型是在哪个数据集上训练的结果呢

Have you provided this code? Where is this module located? Thank you for your help.

当我运行train_cstrack.py进行训练自己的数据集后，生成best_mot_test.pt权重。再将此权重进行测试，发现测试的结果图像上没有显示关联的id，请问这是什么原因？该如何解决呢？万分感谢！

Hi, Table 1 of your paper "Learn to Match: Automatic Matching Network Design for Visual Tracking" provides the comparison of the performance of different challenging attributes on the OTB dataset. Can you provide the position of your codebase where implement this feature?

Hi author, thank you for your excellent work. It is mentioned in the article that Ocean’s DW-XCorr is replaced with concatenation, but the Ocean feature combinationpart uses different dilation, resulting in irregular output sizes, and the final concatenation results cannot be directly added. So I want to know how to deal with this problem?

您好，我有几个小问题需要向您请教一下，西北工业大学在读学生，直博生在读，博士课题就是多目标跟踪。
第一个问题是，我在mot benchmark上注册一直没有回应，注册了很多次，用学校邮箱或者其他邮箱，都一年了，没有任何回应。
第二个问题是，现在在使用很早之前一个师兄注册的账号，但是最近突然发现提交不上去了，能到用户界面，也能load 结果文件，提交的时候也会显示提交成功，界面说是会在几分钟内返回结果，但是没有任何回应，刷新个人用户界面时，显示的提交次数是0，最近两周试了很多次，都是这样。
比较好奇是因为西工大邮箱或者ip问题还是怎么回事儿，其他同方向的伙伴们可以提交和注册吗

Traceback (most recent call last):
File "train_cstrack.py", line 513,
train(hyp, opt, device, tb_writer)
File "train_cstrack.py", line 119, in train
dataloader, dataset = create_dataloader(dataset_root,trainset_paths, imgsz, batch_size, gs, opt, hyp=hyp, augment=True,
File "/home/mes/Projects/SOTS/tracking/../lib/dataset/cstrack.py", line 63, in create_dataloader
dataset = LoadImagesAndLabels(root,path, imgsz, batch_size,
File "/home/mes/Projects/SOTS/tracking/../lib/dataset/cstrack.py", line 617, in init
cache = self.cache_labels(cache_path) # cache
File "/home/mes/Projects/SOTS/tracking/../lib/dataset/cstrack.py", line 776, in cache_labels
torch.save(x, path) # save for next time
File "/home/mes/.conda/envs/CSTrack/lib/python3.8/site-packages/torch/serialization.py", line 361, in save
with _open_file_like(f, 'wb') as opened_file:
File "/home/mes/.conda/envs/CSTrack/lib/python3.8/site-packages/torch/serialization.py", line 229, in _open_file_like
return _open_file(name_or_buffer, mode)
File "/home/mes/.conda/envs/CSTrack/lib/python3.8/site-packages/torch/serialization.py", line 210, in init
super(_open_file, self).init(open(name, mode))
FileNotFoundError: [Errno 2] No such file or directory: './dataset/train.cache'

Thanks for your great work! I noticed that CSTrack use the attention mechanism as one of the important module, but you do not provide the attention map visualization results in your paper. And I'm wondering how to visualize the attention map.

具体代码见：
这一行。
我仔细阅读了数据集这里的代码，发现制作分类的label这里，如果不使用这里aug_apply返回的real_param，制作出的标签似乎脱离了groundtruth，这里是否应该改为crop_bbox, param = aug_apply(Corner(*crop_bbox), param, shape)?如果是我的理解有误，请多多指教！

Hello, I have a question. Can CSTrack track categories other than people? Thanks for the answer!

老师您好,我最近在学习您这个项目中的一些代码,我不是很清楚test_cstrack_panda_post_process.py文件中参数的含义,如果有可能的话,能麻烦您简单解释一下吗?是为了优化MOTA分数吗

您好，感谢您提出的卓越工作！
这里有一些让我存在困惑的结果问题，希望向您请教一下！

我在 Ocean 论文中得到到的VOT2019结果为

offline版本 A 0.590 | R 0.376 | EAO 0.327
online版本 A 0.594 | R 0.316 | EAO 0.350

但您提供的 Google Drive 中的结果只有 Ocean_VOT2019_0.323 文件夹，测试的指标为

A 0.598 | R  0.401 | EAO 0.323

请问 EAO 0.323 的是哪一个版本呢？为什么和论文会有这么大的差异呢？

谢谢！💖

作者您好，关于您的论文我想问您一个问题。您的 Concatenation 算子中，模板帧特征经过roi后是通过copy这组特征得到一个和search同尺寸的特征，对吗

Thank you for your nice recent paper, "One More Check: Making “Fake Background” Be Tracked Again" which shows high improvement in performance compared to the current state of the art in Multi-object tracking. The link on the paper direct to this repository is where we can find the implementation, however, I didn't find its implementation.

Also on the paper " https://arxiv.org/pdf/2104.09441v1.pdf "stated that other information can be found in Supplementary material, but also I didn't find those supplementary materials.

Can you share those supplementary materials? Also, when do you expect to release the code of this paper??

Thank you again for your nice work.

Hello! I have a question about max_index setting in LoadImagesAndLabels

for ds, label_paths in self.label_files_dict.items():
          max_index = -1
          for lp in label_paths:
              lb = np.loadtxt(lp)
              if len(lb) < 1:
                  continue
              if len(lb.shape) < 2:
                  img_max = lb[1]
              else:
                  img_max = np.max(lb[:, 1])
              if img_max > max_index:
                  max_index = img_max
          self.tid_num[ds] = max_index + 1 # here

      last_index = 0
      for i, (k, v) in enumerate(self.tid_num.items()):
                self.tid_start_index[k] = last_index
                last_index += v

when self.tid_num[ds]=max_index+1 is applied to all dataset, index i think that is not used occur.

ex)

Between mot20 and mot16, an index 2733 is not used.

Does you have any other reason why self.tid_num[ds]=max_index+1 is applied to all dataset, not applying only first dataset?

Since these two both cover the whole Ocean method.

Hello, I want to save the output result as a video when I run test_cstrack.py. but I find that vis.plot_tracking is missing on line 89. Would you please tell me which file it is in?

Your job is great and inspired me a lot ( I'm currently working on MOT and have gained some promised result ). I notice that your result on paper ( 70.6 MOTA) and on Mot Server (74.8 MOTA) is different. Could you inform what improve you have done ( i.e change the backbone) and update your work to new version of the paper on arxiv, thanks very much

感谢开源
tutorial中链接给出的yolov5l.pt 是仅包括yolo官方给出的backbone和检测head权重吗？
还是包含了CCN,SAAN这些部分，您自己进行预训练的呢？

When I run setup.py, it won't work. I notice this may be because deformable conv can't be compiled in Cuda 11.1 for RTX3080(can't use Cuda10 there). Have you encountered this problem?

老项目的用的pytorch版本一致，都是1.1，DCN编译依赖torch版本，可否考虑将项目torch版本升级到1.6以上？

#ID_loss
        if i == 0:
            ps_id = id_embeding[indices_id[i]]

In class 'MOTLoss', the code above uses only first value of indices_id.

This means,I think, that
- indices_id[0] has the locations in which an small object exists. (it operates like mask tensor)
- and, the 1/8-scale features are only used to compute the ID Loss. (only small objects are used for ID-Loss)

Did I understand the code in right way?

Thanks for your wonderful work! I find that on MOTChallenge dataset clipping bbox may lead to degraded results. I comment all the clip bbox operations in https://github.com/JudasDie/SOTS/blob/master/lib/dataset/cstrack.py and meet the following training error. Can you help me fix this? Thanks very much. @H11zang

Traceback (most recent call last):
File "train_cstrack.py", line 517, in
train(hyp, opt, device, tb_writer)
File "train_cstrack.py", line 319, in train
loss, loss_items = mot_loss(pred, targets.to(device), model) # scaled by batch_size
File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/opt/tiger/demo/SOTS/tracking/../lib/core/mot/base_trainer.py", line 355, in forward
giou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # giou(prediction, target)
File "/opt/tiger/demo/SOTS/tracking/../lib/core/mot/base_trainer.py", line 67, in bbox_iou
w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1
RuntimeError: CUDA error: device-side assert triggered
/pytorch/aten/src/ATen/native/cuda/IndexKernel.cu:84: operator(): block: [79,0,0], thread: [96,0,0] Assertion index >= -sizes[i] && index < sizes[i] && "index out of bounds" failed.
/pytorch/aten/src/ATen/native/cuda/IndexKernel.cu:84: operator(): block: [79,0,0], thread: [97,0,0] Assertion index >= -sizes[i] && index < sizes[i] && "index out of bounds" failed.

Hi
Really admire the work "One More Check: Making “Fake Background” Be Tracked Again".
Just kindly ask, any plans to release the corresponding code?
Thanks.