search

TexasInstruments / edgeai-yolov5

YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite. Forked from https://ultralytics.com/yolov5
https://github.com/TexasInstruments/edgeai
GNU General Public License v3.0
650 stars 120 forks source link

keypoint prediction error #95

Open sankexin opened 1 year ago

sankexin commented 1 year ago

detect by this weight: image result: image I was beginning to think it was a post-process error,but when I print the pred of detect.py,the y of keypoint of the person is Almost the same size: tensor([[219.79369, 225.33395, 226.03517, 225.67001, 220.89984, 223.56926, 219.32716, 228.19868, 227.60158, 224.10791, 220.28212, 227.59915, 225.78040, 223.75421, 224.11528, 226.45985, 228.28322], [411.84937, 418.37204, 417.93945, 417.42044, 412.42886, 414.72122, 411.45743, 419.45306, 419.69476, 416.37424, 412.08868, 420.05914, 417.24695, 415.01300, 415.85199, 418.92923, 420.43335]]) 640x640 2 persons, Done. (0.047s) so, the newest Pre-training weights are also not correct! could you find find any reason and update zhe correct one? and will you update yolov5(6.0) ?there is no focus. and then I test Yolov5s6_pose_640_ti_lite, there is some bug: yolov5-pose/utils/plots.py, line 118, in plot_skeleton_kpts pos1 = (int(kpts[(sk[0]-1)steps]), int(kpts[(sk[0]-1)steps+1])) IndexError: index 45 is out of bounds for dimension 0 with size 0

sankexin commented 1 year ago

I have solved by this way:

lkpt += kpt_loss_factor((1 - torch.exp(-d/(s(4*sigmas*2)+1e-9)))kpt_mask).mean() # not correct equation

lkpt += kpt_loss_factor((1 - torch.exp(-d/(2(s*sigmas)*2+1e-9)))kpt_mask).mean() lkpt += (self.kptloss(tkpt[i][:, 0::2], pkpt_x, kpt_mask) + self.kptloss(tkpt[i][:, 1::2], pkpt_y, kpt_mask)) / 2 reference: https://github.com/qinggangwu/yolov7-pose_Npoint_Ncla/blob/db5ebf25060f723cce084bdaafd277847ff38d27/utils/loss_Ncla.py

then train it myself.

nomaad42 commented 1 year ago

I have solved by this way:

lkpt += kpt_loss_factor((1 - torch.exp(-d/(s(4*sigmas*2)+1e-9)))kpt_mask).mean() # not correct equation

lkpt += kpt_loss_factor((1 - torch.exp(-d/(2(s*sigmas)*2+1e-9)))kpt_mask).mean() lkpt += (self.kptloss(tkpt[i][:, 0::2], pkpt_x, kpt_mask) + self.kptloss(tkpt[i][:, 1::2], pkpt_y, kpt_mask)) / 2 reference: https://github.com/qinggangwu/yolov7-pose_Npoint_Ncla/blob/db5ebf25060f723cce084bdaafd277847ff38d27/utils/loss_Ncla.py

then train it myself.

Hi! @sankexin

Are you using both of these equations?

lkpt += kpt_loss_factor*((1 - torch.exp(-d/(2*(s*sigmas)**2+1e-9)))*kpt_mask).mean() lkpt += (self.kptloss(tkpt[i][:, 0::2], pkpt_x, kpt_mask) + self.kptloss(tkpt[i][:, 1::2], pkpt_y, kpt_mask)) / 2

Or just the one at the bottom?

sankexin commented 1 year ago

这是来自QQ邮箱的假期自动回复邮件。   您好，我最近正在休假中，无法亲自回复您的邮件。我将在假期结束后，尽快给您回复。

Song-Jie-ing commented 6 months ago

I have solved by this way:

lkpt += kpt_loss_factor((1 - torch.exp(-d/(s(4*sigmas*2)+1e-9)))kpt_mask).mean() # not correct equation

lkpt += kpt_loss_factor((1 - torch.exp(-d/(2(s*sigmas)*2+1e-9)))kpt_mask).mean() lkpt += (self.kptloss(tkpt[i][:, 0::2], pkpt_x, kpt_mask) + self.kptloss(tkpt[i][:, 1::2], pkpt_y, kpt_mask)) / 2 reference: https://github.com/qinggangwu/yolov7-pose_Npoint_Ncla/blob/db5ebf25060f723cce084bdaafd277847ff38d27/utils/loss_Ncla.py then train it myself.

Hi! @sankexin

Are you using both of these equations?

lkpt += kpt_loss_factor*((1 - torch.exp(-d/(2*(s*sigmas)**2+1e-9)))*kpt_mask).mean() lkpt += (self.kptloss(tkpt[i][:, 0::2], pkpt_x, kpt_mask) + self.kptloss(tkpt[i][:, 1::2], pkpt_y, kpt_mask)) / 2

Or just the one at the bottom?

hello! Are you work out this problem?

Song-Jie-ing commented 6 months ago

lkpt += kptloss

这是来自QQ邮箱的假期自动回复邮件。   您好，我最近正在休假中，无法亲自回复您的邮件。我将在假期结束后，尽快给您回复。

你好，请问这个问题你解决了吗