ValueError: Anchors should be Tuple[Tuple[int]] because each feature map could potentially have different sizes and aspect ratios. There needs to be a match between the number of feature maps passed and the number of sizes / aspect ratios specified.

when I try to run the tracker on a CroHD dataset(training set) and **set the "use_public" option in the 'det_cfg' to 'False',**the ValueError said that "Anchors should be Tuple[Tuple[int]] because each feature map could potentially have different sizes and aspect ratios. " I noticed that in the obj_detect.py,if det_cfg['median_anchor']:the program can choose different benchmark to import different anchors,so which benchmark I should set in the det_cfg when I want to run the tracker on a CroHD dataset?And I noticed that the benchmark should be associated with the anchor.py,too. Another question is,if i set the "use_public" option in the 'det_cfg' to 'True',does it mean that I will use the data in det.txt which the dataset offers and the detector will not be used to detect the pedestrians or output anything.

File "/home/project/HeadHunter-master/head_detection/models/fast_rcnn.py", line 496, in forward boxes, scores = self.filter_proposals(proposals, objectness, images.image_sizes, num_anchors_per_level) File "/home/project/HeadHunter-master/head_detection/models/fast_rcnn.py", line 458, in filter_proposals keep = keep[:self.post_nms_top_n] TypeError: slice indices must be integers or None or have an index method

I run test.py. args.pretrained_model is from this link. I meet this error: RuntimeError: Error(s) in loading state_dict for FasterRCNN: Unexpected key(s) in state_dict: "backbone.ssh1.branch1.0.weight", "backbone.ssh1.branch1.1.weight", "backbone.ssh1.branch1.1.bias", "backbone.ssh1.branch1.1.running_mean", "backbone.ssh1.branch1.1.running_var", "backbone.ssh1.branch1.1.num_batches_tracked", "backbone.ssh1.branch2a.0.weight", "backbone.ssh1.branch2a.1.weight", "backbone.ssh1.branch2a.1.bias", "backbone.ssh1.branch2a.1.running_mean", "backbone.ssh1.branch2a.1.running_var", "backbone.ssh1.branch2a.1.num_batches_tracked", "backbone.ssh1.branch2b.0.weight", "backbone.ssh1.branch2b.1.weight", "backbone.ssh1.branch2b.1.bias", "backbone.ssh1.branch2b.1.running_mean", "backbone.ssh1.branch2b.1.running_var", "backbone.ssh1.branch2b.1.num_batches_tracked", "backbone.ssh1.branch2c.0.weight", "backbone.ssh1.branch2c.1.weight", "backbone.ssh1.branch2c.1.bias", "backbone.ssh1.branch2c.1.running_mean", "backbone.ssh1.branch2c.1.running_var", "backbone.ssh1.branch2c.1.num_batches_tracked", "backbone.ssh1.ssh_1.weight", "backbone.ssh1.ssh_1.bias", "backbone.ssh1.ssh_dimred.weight", "backbone.ssh1.ssh_dimred.bias", "backbone.ssh1.ssh_2.weight", "backbone.ssh1.ssh_2.bias", "backbone.ssh1.ssh_3a.weight", "backbone.ssh1.ssh_3a.bias", "backbone.ssh1.ssh_3b.weight", "backbone.ssh1.ssh_3b.bias", "backbone.ssh1.ssh_final.0.weight", "backbone.ssh1.ssh_final.1.weight", "backbone.ssh1.ssh_final.1.bias", "backbone.ssh1.ssh_final.1.running_mean", "backbone.ssh1.ssh_final.1.running_var", "backbone.ssh1.ssh_final.1.num_batches_tracked", "backbone.ssh2.branch1.0.weight", "backbone.ssh2.branch1.1.weight", "backbone.ssh2.branch1.1.bias", "backbone.ssh2.branch1.1.running_mean", "backbone.ssh2.branch1.1.running_var", "backbone.ssh2.branch1.1.num_batches_tracked", "backbone.ssh2.branch2a.0.weight", "backbone.ssh2.branch2a.1.weight", "backbone.ssh2.branch2a.1.bias", "backbone.ssh2.branch2a.1.running_mean", "backbone.ssh2.branch2a.1.running_var", "backbone.ssh2.branch2a.1.num_batches_tracked", "backbone.ssh2.branch2b.0.weight", "backbone.ssh2.branch2b.1.weight", "backbone.ssh2.branch2b.1.bias", "backbone.ssh2.branch2b.1.running_mean", "backbone.ssh2.branch2b.1.running_var", "backbone.ssh2.branch2b.1.num_batches_tracked", "backbone.ssh2.branch2c.0.weight", "backbone.ssh2.branch2c.1.weight", "backbone.ssh2.branch2c.1.bias", "backbone.ssh2.branch2c.1.running_mean", "backbone.ssh2.branch2c.1.running_var", "backbone.ssh2.branch2c.1.num_batches_tracked", "backbone.ssh2.ssh_1.weight", "backbone.ssh2.ssh_1.bias", "backbone.ssh2.ssh_dimred.weight", "backbone.ssh2.ssh_dimred.bias", "backbone.ssh2.ssh_2.weight", "backbone.ssh2.ssh_2.bias", "backbone.ssh2.ssh_3a.weight", "backbone.ssh2.ssh_3a.bias", "backbone.ssh2.ssh_3b.weight", "backbone.ssh2.ssh_3b.bias", "backbone.ssh2.ssh_final.0.weight", "backbone.ssh2.ssh_final.1.weight", "backbone.ssh2.ssh_final.1.bias", "backbone.ssh2.ssh_final.1.running_mean", "backbone.ssh2.ssh_final.1.running_var", "backbone.ssh2.ssh_final.1.num_batches_tracked", "backbone.ssh3.branch1.0.weight", "backbone.ssh3.branch1.1.weight", "backbone.ssh3.branch1.1.bias", "backbone.ssh3.branch1.1.running_mean", "backbone.ssh3.branch1.1.running_var", "backbone.ssh3.branch1.1.num_batches_tracked", "backbone.ssh3.branch2a.0.weight", "backbone.ssh3.branch2a.1.weight", "backbone.ssh3.branch2a.1.bias", "backbone.ssh3.branch2a.1.running_mean", "backbone.ssh3.branch2a.1.running_var", "backbone.ssh3.branch2a.1.num_batches_tracked", "backbone.ssh3.branch2b.0.weight", "backbone.ssh3.branch2b.1.weight", "backbone.ssh3.branch2b.1.bias", "backbone.ssh3.branch2b.1.running_mean", "backbone.ssh3.branch2b.1.running_var", "backbone.ssh3.branch2b.1.num_batches_tracked", "backbone.ssh3.branch2c.0.weight", "backbone.ssh3.branch2c.1.weight", "backbone.ssh3.branch2c.1.bias", "backbone.ssh3.branch2c.1.running_mean", "backbone.ssh3.branch2c.1.running_var", "backbone.ssh3.branch2c.1.num_batches_tracked", "backbone.ssh3.ssh_1.weight", "backbone.ssh3.ssh_1.bias", "backbone.ssh3.ssh_dimred.weight", "backbone.ssh3.ssh_dimred.bias", "backbone.ssh3.ssh_2.weight", "backbone.ssh3.ssh_2.bias", "backbone.ssh3.ssh_3a.weight", "backbone.ssh3.ssh_3a.bias", "backbone.ssh3.ssh_3b.weight", "backbone.ssh3.ssh_3b.bias", "backbone.ssh3.ssh_final.0.weight", "backbone.ssh3.ssh_final.1.weight", "backbone.ssh3.ssh_final.1.bias", "backbone.ssh3.ssh_final.1.running_mean", "backbone.ssh3.ssh_final.1.running_var", "backbone.ssh3.ssh_final.1.num_batches_tracked", "backbone.ssh4.branch1.0.weight", "backbone.ssh4.branch1.1.weight", "backbone.ssh4.branch1.1.bias", "backbone.ssh4.branch1.1.running_mean", "backbone.ssh4.branch1.1.running_var", "backbone.ssh4.branch1.1.num_batches_tracked", "backbone.ssh4.branch2a.0.weight", "backbone.ssh4.branch2a.1.weight", "backbone.ssh4.branch2a.1.bias", "backbone.ssh4.branch2a.1.running_mean", "backbone.ssh4.branch2a.1.running_var", "backbone.ssh4.branch2a.1.num_batches_tracked", "backbone.ssh4.branch2b.0.weight", "backbone.ssh4.branch2b.1.weight", "backbone.ssh4.branch2b.1.bias", "backbone.ssh4.branch2b.1.running_mean", "backbone.ssh4.branch2b.1.running_var", "backbone.ssh4.branch2b.1.num_batches_tracked", "backbone.ssh4.branch2c.0.weight", "backbone.ssh4.branch2c.1.weight", "backbone.ssh4.branch2c.1.bias", "backbone.ssh4.branch2c.1.running_mean", "backbone.ssh4.branch2c.1.running_var", "backbone.ssh4.branch2c.1.num_batches_tracked", "backbone.ssh4.ssh_1.weight", "backbone.ssh4.ssh_1.bias", "backbone.ssh4.ssh_dimred.weight", "backbone.ssh4.ssh_dimred.bias", "backbone.ssh4.ssh_2.weight", "backbone.ssh4.ssh_2.bias", "backbone.ssh4.ssh_3a.weight", "backbone.ssh4.ssh_3a.bias", "backbone.ssh4.ssh_3b.weight", "backbone.ssh4.ssh_3b.bias", "backbone.ssh4.ssh_final.0.weight", "backbone.ssh4.ssh_final.1.weight", "backbone.ssh4.ssh_final.1.bias", "backbone.ssh4.ssh_final.1.running_mean", "backbone.ssh4.ssh_final.1.running_var", "backbone.ssh4.ssh_final.1.num_batches_tracked". what should i do?

anchors = self.grid_anchors(grid_sizes, strides) File "/home/haojie/下载/ENTER/envs/head_detection2/lib/python3.8/site-packages/torchvision/models/detection/anchor_utils.py", line 103, in grid_anchors raise ValueError("Anchors should be Tuple[Tuple[int]] because each feature " ValueError: Anchors should be Tuple[Tuple[int]] because each feature map could potentially have different sizes and aspect ratios. There needs to be a match between the number of feature maps passed and the number of sizes / aspect ratios specified. Killing subprocess 1055

How much GPU memory is recommanded when training the model using ScutHead dataset? I've found a memory insufficience error when trying to train the model on a idel GPU with 4041MB memory and batch size 1. Besides, there is some minor bugs in the "head_detection/data/create_scuthead.py" file, as args.dset_path is defined but args.dset_dir is used (and the same bugs to the arg.save_path and args.out_path).

Thanks for providing the code to your paper. I managed to get started, even though my gpu only has 8 GB. I just added A.SmallestMaxSize(max_size=400, p=1.0), to the transforms in dataset.py.

Everything else is just used as is on the ScutHead-Dataset (and correct versions of torch=1.6.0 Now I get the error, directly on the first iteration: Loss is nan, stopping training

I suppose this is because in fast_rcnn.py CustomRoIHead.forward the passed features are filled with nan Also setting ohem=soft_nms=upscale_rpn=False and thereby using torchvisions RoiHead doesn't help. Did you experience something like this during training?

Hi @Sentient07,

I'm confused about some results in your "CVPR21 paper: Tracking Pedestrian Heads in Dense Crowd" cause some details are missing in the paper. Hope you can help me sort them out. Thanks!

1. The test set of SCUT-HEAD dataset consists of two parts: Part-A and Part-B. I find that other methods compared in Table 2 report respective results on both parts in their papers. So, which part did you use to run the evaluation to obtain the results in Table 2? Or is it the fact that the results in Table 2 are obtained by evaluating methods on the combination of both parts? Since I couldn't find the exact results of other compared methods in Table 2 in the corresponding cited papers, I'd like to know how did you obtain the results of these methods?

2. Table 4 shows the tracking results of your proposed tracker HeadHunter-T and other state-of-the-art trackers on the test set of CroHD. However, the results of HeadHunter-T in Table 4 are quite different from the results shown on the website of the MOT benchmark. Can you tell me what's the difference between these two results?

Looking forward to your reply. Thanks!

Two questions： 1. Is this model suitable for low-density pedestrian tracking? 2. The groudTruth file is loaded in a different format than the readme file

Two questions： 1. Can I use this model to train our own data set？If so, what are the differences from the three data sets mentioned by the author. 2. The groudTruth file is loaded in a different format than the readme file

Hi, I am very interested in your great work, I try to reproduce your work from scratch. But I cannot find anywhere to download CroHD dataset. Could you please give me a way to download it?

Why are the results published on the croHD test set in your paper different from those published on the mot challeng website. The MOTA published in the paper is 63.6, compared to 57.8 on the mot challenge website

Has anyone been able to run the program properly? I followed the instruction in the repo, but haven't got any luck yet. I am confused by the mismatch of cuda versions between the one indicated in this repo and the one actually used by the author, and the extra packages I need to install to stop receiving errors. Below are the packages I installed aside from the one in the requirements, but I still couldn't get any output.

pip install pyyaml
pip install ipykernel
pip install albumentations==0.4.6
pip install scipy==1.1.0
pip install pycocotools
pip install munkres
pip install scikit-image==0.16.2

Just to put a context: I was asked to find a paper and reproduce some results from scratch (it weights the 50% of the subject), I've my deadline around the 10 of June of 2022.

While rewriting the detection network (in order to fully understand the paper) I found strange the CPM part and I would like to ask for advice.

Papers Text

The paper says:

with Context Sensitive feature extractor followed by series of transpose convolutions to enhance spatial resolution of feature maps.

and

we augmented on top of each individual FPNs, a Context-sensitive Prediction Module (CPM) [63]. This contextual module consists of 4 Inception-ResNet-A blocks [62] with 128 and 256 filters for 3 × 3 convolution and 1024 filters for 1 × 1 convolution.

The reference 63 says:

We design the Context-sensitive Predict Module (CPM), see Fig. 3(b), in which we replace the convolution layers of context module in SSH by the residual-free prediction module of DSSD.

Issues

From the previous cites, I understand the CPM as a SSH with different convolution operations. But your Figure 4 (from the paper) and your code shows a channel expansion which seems like the prediction module of DSSD (a kind of simplified Inception) followed by a standard SSH.

I did not find any Inception-ResNet-A blocks.

Additionally, I did not find the transpose convolutions part.

Sorry for the inconvenience, I just want to make sure I don't miss any detail and have it done correctly as soon as possible...

Hello, If I try to run test.py the pretrained_model you provided on CroHD, I am facing a problem with Anchors:

python test.py --test_dataset CroHD/test/HT21-11/img1 --plot_folder outputs --outfile outputs --pretrained_model FT_R50_epoch_24.pth --context cpm

Output, with the Traceback:

256
FT_R50_epoch_24.pth
0it [00:00, ?it/s]/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torch/nn/functional.py:3502: UserWarning: The default behavior for interpolate/upsample with float scale_factor changed in 1.6.0 to align with other frameworks/libraries, and now uses scale_factor directly, instead of relying on the computed output size. If you wish to restore the old behavior, please set recompute_scale_factor=True. See the documentation of nn.Upsample for details. 
  warnings.warn(
0it [00:02, ?it/s]
Traceback (most recent call last):
  File "test.py", line 176, in <module>
    test()
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "test.py", line 165, in test
    outputs = model(images)
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torchvision/models/detection/generalized_rcnn.py", line 97, in forward
    proposals, proposal_losses = self.rpn(images, features, targets)
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torchvision/models/detection/rpn.py", line 345, in forward
    anchors = self.anchor_generator(images, features)
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torchvision/models/detection/anchor_utils.py", line 150, in forward
    anchors_over_all_feature_maps = self.cached_grid_anchors(grid_sizes, strides)
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torchvision/models/detection/anchor_utils.py", line 139, in cached_grid_anchors
    anchors = self.grid_anchors(grid_sizes, strides)
  File "/mnt/sdb/anaconda3/envs/headhunter-TT/lib/python3.8/site-packages/torchvision/models/detection/anchor_utils.py", line 103, in grid_anchors
    raise ValueError("Anchors should be Tuple[Tuple[int]] because each feature "
ValueError: Anchors should be Tuple[Tuple[int]] because each feature map could potentially have different sizes and aspect ratios. There needs to be a match between the number of feature maps passed and the number of sizes / aspect ratios specified.

These are the contents of my virtual environment:

name: headhunter-TT
channels:
  - defaults
dependencies:
  - _libgcc_mutex=0.1=main
  - ca-certificates=2021.4.13=h06a4308_1
  - certifi=2020.12.5=py38h06a4308_0
  - ld_impl_linux-64=2.33.1=h53a641e_7
  - libffi=3.3=he6710b0_2
  - libgcc-ng=9.1.0=hdf63c60_0
  - libstdcxx-ng=9.1.0=hdf63c60_0
  - ncurses=6.2=he6710b0_1
  - openssl=1.1.1k=h27cfd23_0
  - pip=21.1.1=py38h06a4308_0
  - python=3.8.10=hdb3f193_7
  - readline=8.1=h27cfd23_0
  - setuptools=52.0.0=py38h06a4308_0
  - sqlite=3.35.4=hdfb4753_0
  - tk=8.6.10=hbc83047_0
  - wheel=0.36.2=pyhd3eb1b0_0
  - xz=5.2.5=h7b6447c_0
  - zlib=1.2.11=h7b6447c_3
  - pip:
    - chardet==4.0.0
    - cycler==0.10.0
    - decorator==4.4.2
    - h5py==3.2.1
    - idna==2.10
    - imageio==2.9.0
    - kiwisolver==1.3.1
    - matplotlib==3.4.2
    - networkx==2.5.1
    - numpy==1.20.3
    - ordered-set==4.0.2
    - pillow==8.2.0
    - plyfile==0.7.4
    - pyparsing==2.4.7
    - python-dateutil==2.8.1
    - pywavelets==1.1.1
    - requests==2.25.1
    - scikit-image==0.18.1
    - scipy==1.6.3
    - six==1.16.0
    - tifffile==2021.4.8
    - torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
    - torchmeta==1.7.0
    - torchvision==0.9.1
    - tqdm==4.61.0
    - trimesh==3.9.19
    - typing-extensions==3.10.0.0
    - urllib3==1.26.5
    - munkres
    - albumentations==0.5.2
    - pyyaml 

If I try to downgrade torch to 1.6.0 and torchvision to 0.7.0, I run through the following error:

RuntimeError: CUDA error: no kernel image is available for execution on the device

Moreover, I get this warning message if I use torch 1.6.0, when I try to get a device info via: torch.cuda.get_device_name(0)

NVIDIA GeForce RTX 3xxx with CUDA capability sm_86 is not compatible with the current PyTorch installation. The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_70.