taeinkwon / h2odataset Goto Github PK

Hi. It seems that I cannot access the dataset by either running the download_script.py file or accessing the h2odataset.ethz.ch/data/ website. When I run the download_script.py file with my username and pswd, it shows that:

raise HTTPError(http_error_msg, response=self)
requests.exceptions.HTTPError: 403 Client Error: Forbidden for url: https://h2odataset.ethz.ch/data/dataset/object.zip

While when I directly check the h2odataset.ethz.ch/data/ website, the List of Files is empty.

Therefore, I wonder whether you can help me with this issue.

Thanks!

Best regards,
Frances

Hi! Thanks for your great work. I try to test my model on test split (as issue #11 ), which is to submit the results to https://codalab.lisn.upsaclay.fr/competitions/4820 . I send the request to participate in this challenge, but it seems that the decision of approval is still pending.

https://github.com/taeinkwon/h2odataset#news says that

Our 1st phase of the challenges will be closed at October 1st, 2022.

So is the CodaLab page still on working now? Is there any other way to get the action labels of the test split?

Hi, this paper is a good job for the further exploration of human objects interacting.
By the way, I would like to know when will you release your dataset.
Thanks.

I would like to know if you have tested your model on other datasets - for example, the epic kitchens dataset

Thank you for sharing this wonderful dataset.

I found a very minor naming error.
download_scirpt.py --> download_script.py

@taeinkwon Hi， I find h2o_CASA.tar in the downlink has subject folders 1-10， which include pose，mano param，etc. At the same time, subjectX_v1_1.tar.gz 1-4 are also in the download page. What's the difference between them？

Hello,

First of all, I would like to congratulate you on this great work.

Could you please indicate the download link for the data set?

Thank you

Hi, thanks for the great work. I try to reproduce the recognization results of SlowFast using pyslowfast. Concretely, I build the dataset with the trimed action clips (i.e. 569 training samples) from the cam4(rgb256) and modify the dataset based on charades.py. The model(SLOWFAST_4x16_R50) is trained from scratch. However, I get about 42% top1 acc on val set for about 100 epochs. Did I miss any details?

Hi,

Thanks for your great job! I have two questions about the labels provided in this dataset.

i) What do the values in hand poses mean? For example, -0.03166291117668152 and 0.11372638493776321.
ii) I see that the hand pose labels are available for the testing set. Do I misunderstand it?

For the challenge host in ECCV2022, can I use other pre-trained models like k400 or epic-kitchen？
Also, I can only use the training set rather than training+validation. Am I right?

Best

Hello.
For Fig 7 (main text) and Tab S5 (supplementary), it seems that the evaluation is to compare the estimation and gt under the camera coordinate system. However, the purely RGB-based hand estimation may suffer from scale ambiguity, therefore the error may be largely influenced by an inaccurate global depth estimation, especially considering that the test subject (Subject 4) does not appear in the training data.
Therefore, I wonder whether the number/curve of hand pose error regarding the wrist-aligned coordinate system for the proposed baseline method can be provided, as this is often adopted by many works that focus on RGB-based hand pose estimation.

Thanks for your valuable help!

Best regards,
Frances

Hi @taeinkwon ,

Thanks for your work, Would you please present more details about your dataset?

1. How much does your dataset would take (Totally) and (images) and (depth images)?
2. Would you please provide visualisation scripts for your dataset?

Hi, I wonder whether the action labels of the testing split are available. I can find the action label for the training and validation set, however, I fail to find that for sequences of testing sets.

Thanks for your help!

Best regards,
Frances

Service Unavailable
The server is temporarily unable to service your request due to maintenance downtime or capacity problems. Please try again later.

Apache/2.4.6 (Red Hat Enterprise Linux) Server at h2odataset.ethz.ch Port 443

Could you tell me about the problem I met when downloading the dataset? Thanks a lot !

Forbidden
You don't have permission to access this resource.
Apache/2.4.37 (Red Hat Enterprise Linux) Server at h2odataset.ethz.ch Port 443

Hi,

I wanted to test the code on other images. May I know when you will release the code?

Thank You

Hi!
Thanks for your great work!It's very excellent.

I do some code to show mano model by myself.it seems mano model not aligned with xyz_cam.My question is whether this is normal

Here are my results

The code mainly comes from freihand
`

def show_opendr_render(mano_model, xyz_cam, cam_intrinsics, img_path, root_id=0, project_func=None):

def calc_global_t():
    V = mano_model
    J_regressor = mano_model.J_regressor
    Jtr_x = ch.MatVecMult(J_regressor, V[:, 0])
    Jtr_y = ch.MatVecMult(J_regressor, V[:, 1])
    Jtr_z = ch.MatVecMult(J_regressor, V[:, 2])
    Jtr = ch.vstack([Jtr_x, Jtr_y, Jtr_z]).T
    xyz_root = Jtr[0, :]
    return xyz_cam[root_id] - xyz_root

def show_hand(image_show, uv):
    for uv_index, (x, y) in enumerate(uv):
        if uv_index >= 1 and uv_index < 5:
            cv2.circle(image_show, (int(x), int(y)), radius=6,
                       color=(0, 0, 255), thickness=-1)
        if uv_index >= 5 and uv_index < 9:
            cv2.circle(image_show, (int(x), int(y)), radius=6,
                       color=(255, 0, 0), thickness=-1)
        if uv_index >= 9 and uv_index < 13:
            cv2.circle(image_show, (int(x), int(y)), radius=6,
                       color=(255, 255, 0), thickness=-1)
        if uv_index >= 13 and uv_index < 17:
            cv2.circle(image_show, (int(x), int(y)), radius=6,
                       color=(0, 255, 255), thickness=-1)
        if uv_index >= 17 and uv_index < 21:
            cv2.circle(image_show, (int(x), int(y)), radius=6,
                       color=(0, 128, 128), thickness=-1)
        if uv_index <= 0:
            cv2.circle(image_show, (int(x), int(y)), radius=6,
                       color=(128, 128, 128), thickness=-1)

# Create OpenDR renderer
rn = ColoredRenderer()

# Camera setting
M = np.eye(4)
R = M[:3, :3]
ax, angle = transforms3d.axangles.mat2axangle(R)
rt = ax * angle
rt = rt.flatten()
t = M[:3, 3]
dist = np.zeros(5).flatten()
pp = np.array([cam_intrinsics[0, 2], cam_intrinsics[1, 2]])
f = np.array([cam_intrinsics[0, 0], cam_intrinsics[1, 1]])

image_show = cv2.imread(str(img_path))
w, h = image_show.shape[1], image_show.shape[0]

# Assign attributes to renderer
rn.camera = ProjectPoints(rt=rt,
                          t=t,  # camera translation
                          f=f,  # focal lengths
                          c=pp,  # camera center (principal point)
                          k=dist)  # OpenCv distortion params
rn.frustum = {'near': 0.1, 'far': 5., 'width': w, 'height': h}

V = mano_model + calc_global_t()
F = mano_model.f
rn.set(v=V,
       f=F,
       bgcolor=np.zeros(3))

colors = np.ones_like(V)

# Construct point light sources
rn.vc = LambertianPointLight(f=F,
                             v=V,
                             num_verts=V.shape[0],
                             light_pos=np.array([-1000, -1000, -2000]),
                             vc=0.8 * colors,
                             light_color=np.array([1., 1., 1.]))

rn.vc += LambertianPointLight(f=F,
                              v=V,
                              num_verts=V.shape[0],
                              light_pos=np.array([1000, 1000, -2000]),
                              vc=0.25 * colors,
                              light_color=np.array([1., 1., 1.]))

rn.vc += LambertianPointLight(f=F,
                              v=V,
                              num_verts=V.shape[0],
                              light_pos=np.array([2000, 2000, 2000]),
                              vc=0.1 * colors,
                              light_color=np.array([1., 1., 1.]))

rn.vc += LambertianPointLight(f=F,
                              v=V,
                              num_verts=V.shape[0],
                              light_pos=np.array([-2000, -2000, 2000]),
                              vc=0.1 * colors,
                              light_color=np.array([1., 1., 1.]))


msk_rendered = (np.array(rn.r) * 255).astype(np.uint8)

if project_func is not None:
    UV = project_func(xyz_cam, cam_intrinsics)
    show_hand(image_show, UV)
    show_hand(msk_rendered, UV)

render_image = np.concatenate([image_show, msk_rendered], axis=1)
cv2.imwrite("../a.jpg",render_image)

if render_image.shape[1] > 1000:
    render_image = cv2.resize(render_image, (render_image.shape[1]//2, render_image.shape[0]//2))
cv2.imshow("render", render_image)
cv2.waitKey(1000)
# cv2.destroyAllWindows()

`

Hi, thanks for sharing such a high-quality dataset!
I was wondering how to understand the numbers of the depth images. From here in your visualization code, a depth image is read as an uint16 matrix, in which the values range from 0 to 3186 (or so).

Could you instruct me on how to transform them into values in meters?

how to download this dataset？