Some questions about get the results provided by your paper in Table 4(supervised with CR))

[julyiii]

Thank you for your excellent work! There some diffculties that I can't get the same or even closed results provided by your paper in Table 4(supervised with CR).And I find the hyper-parameters in the code is not configed as your paper.Do your mand providing your hyper-parameters to help me ? Thank you. Here my logs.

====== Options ====== HO3D_root: /datassd/huanyao/semi-hand/Semi-Hand-Object-master/assets/data/ho3d_v2 blocks: 1 channels: 256 epochs: 60 evaluate: False host_folder: /datassd/huanyao/semi-hand/Semi-Hand-Object-master/host_folder inp_res: 512 lambda_joints2d: 100.0 lambda_objects: 500.0 lr: 0.0001 lr_decay_gamma: 0.7 lr_decay_step: 10 mano_lambda_joints3d: 10000.0 mano_lambda_manopose: 10 mano_lambda_manoshape: 0.1 mano_lambda_regulpose: 1 mano_lambda_regulshape: 100.0 mano_lambda_verts3d: 10000.0 mano_neurons: [1024, 512] mano_root: assets/mano_models manual_seed: 0 momentum: 0.9 network: honet_transformer obj_model_root: assets/object_models resume: /datassd/huanyao/semi-hand/Semi-Hand-Object-master/model.pth save_results: True snapshot: 5 stacks: 1 test_batch: 16 test_freq: 10 train_batch: 24 transformer_depth: 1 transformer_head: 1 use_cuda: 1 weight_decay: 0.0005 workers: 16

[stevenlsw]

Hi, could you provide me with your training results? The network architecture is the same as provided in the checkpoint.

On Sun, Jul 31, 2022 at 7:54 PM julyiii @.***> wrote:

Thank you for your excellent work! There some diffculties that I can't get the same or even closed results provided by your paper in Table 4(supervised with CR).And I find the hyper-parameters in the code is not configed as your paper.Do your mand providing your hyper-parameters to help me ? Thank you. Here my logs.

====== Options ====== HO3D_root: /datassd/huanyao/semi-hand/Semi-Hand-Object-master/assets/data/ho3d_v2 blocks: 1 channels: 256 epochs: 60 evaluate: False host_folder: /datassd/huanyao/semi-hand/Semi-Hand-Object-master/host_folder inp_res: 512 lambda_joints2d: 100.0 lambda_objects: 500.0 lr: 0.0001 lr_decay_gamma: 0.7 lr_decay_step: 10 mano_lambda_joints3d: 10000.0 mano_lambda_manopose: 10 mano_lambda_manoshape: 0.1 mano_lambda_regulpose: 1 mano_lambda_regulshape: 100.0 mano_lambda_verts3d: 10000.0 mano_neurons: [1024, 512] mano_root: assets/mano_models manual_seed: 0 momentum: 0.9 network: honet_transformer obj_model_root: assets/object_models resume: /datassd/huanyao/semi-hand/Semi-Hand-Object-master/model.pth save_results: True snapshot: 5 stacks: 1 test_batch: 16 test_freq: 10 train_batch: 24 transformer_depth: 1 transformer_head: 1 use_cuda: 1 weight_decay: 0.0005 workers: 16

— Reply to this email directly, view it on GitHub https://github.com/stevenlsw/Semi-Hand-Object/issues/11, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFQRNGMW7YMHFIOMGZDMHRDVW4N25ANCNFSM55FZ4Z4A . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[julyiii]

This is th result of 60 epoch.

This is th result of 50 epoch. REP-5 {'021_bleach_cleanser': 0.022821576763485476, '006_mustard_bottle': 0.0348314606741573, '010_potted_meat_can': 0.018075801749271137} ADD-10 {'021_bleach_cleanser': 0.6072910491997628, '006_mustard_bottle': 0.6146067415730337, '010_potted_meat_can': 0.2043731778425656} This is th result of 60 epoch. REP-5 {'021_bleach_cleanser': 0.024303497332542976, '006_mustard_bottle': 0.0853932584269663, '010_potted_meat_can': 0.01661807580174927} ADD-10 {'021_bleach_cleanser': 0.5145228215767634, '006_mustard_bottle': 0.6426966292134831, '010_potted_meat_can': 0.1932944606413994}

[stevenlsw]

Hi July,

I don't think the model you trained is well behaved (the performance is quite low), but the training process looks correct to me. Currently I am not sure what the problem is. Could you please train with more epochs and see the result improved.

On Mon, Aug 1, 2022 at 12:09 AM julyiii @.***> wrote:

This is th result of 60 epoch. [image: 屏幕截图 2022-08-01 130320] https://user-images.githubusercontent.com/95080276/182076221-9de76b69-2bfd-428a-81ee-3df7269222db.png

This is th result of 50 epoch. REP-5 {'021_bleach_cleanser': 0.022821576763485476, '006_mustard_bottle': 0.0348314606741573, '010_potted_meat_can': 0.018075801749271137} ADD-10 {'021_bleach_cleanser': 0.6072910491997628, '006_mustard_bottle': 0.6146067415730337, '010_potted_meat_can': 0.2043731778425656} This is th result of 60 epoch. REP-5 {'021_bleach_cleanser': 0.024303497332542976, '006_mustard_bottle': 0.0853932584269663, '010_potted_meat_can': 0.01661807580174927} ADD-10 {'021_bleach_cleanser': 0.5145228215767634, '006_mustard_bottle': 0.6426966292134831, '010_potted_meat_can': 0.1932944606413994}

— Reply to this email directly, view it on GitHub https://github.com/stevenlsw/Semi-Hand-Object/issues/11#issuecomment-1200713971, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFQRNGKJ2QAS7GHBZQ2PI23VW5LYLANCNFSM55FZ4Z4A . You are receiving this because you commented.Message ID: @.***>

[julyiii]

Okk, I will try to train with more epoches to find if the result will be better.Thank you for your reply!

[karta2155802]

Hi, thanks for your remarkable work. I met the similar difficulty as @julyiii. The evaluation metrics didn't perform well. The object ADD seems ok, but the hand estimation is not quite well. Here's my logs at 50 epochs. I only changed the train_batch and workers to fit my hardware settings.

====== Options ======
HO3D_root: ../HandOccNet/data/HO3D/data
blocks: 1
channels: 256
epochs: 60
evaluate: False
host_folder: official
inp_res: 512
lambda_joints2d: 100.0
lambda_objects: 500.0
lr: 0.0001
lr_decay_gamma: 0.7
lr_decay_step: 10
mano_lambda_joints3d: 10000.0
mano_lambda_manopose: 10
mano_lambda_manoshape: 0.1
mano_lambda_regulpose: 1
mano_lambda_regulshape: 100.0
mano_lambda_verts3d: 10000.0
mano_neurons: [1024, 512]
mano_root: assets/mano_models
manual_seed: 0
momentum: 0.9
network: honet_transformer
obj_model_root: assets/object_models
resume: None
save_results: True
snapshot: 1
stacks: 1
test_batch: 16
test_freq: 10
train_batch: 8
transformer_depth: 1
transformer_head: 1
use_cuda: 1
weight_decay: 0.0005
workers: 12
=====================

Here's the results at 50 epochs.

Evaluation 3D KP results:
auc=0.334, mean_kp3d_avg=4.68 cm
Evaluation 3D KP PROCRUSTES ALIGNED results:
auc=0.767, mean_kp3d_avg=1.17 cm
Evaluation 3D KP SCALE-TRANSLATION ALIGNED results:
auc=0.344, mean_kp3d_avg=4.66 cm

Evaluation 3D MESH results:
auc=0.344, mean_kp3d_avg=4.52 cm
Evaluation 3D MESH ALIGNED results:
auc=0.778, mean_kp3d_avg=1.11 cm

F-scores
F@5.0mm = 0.164     F_aligned@5.0mm = 0.466
F@15.0mm = 0.534    F_aligned@15.0mm = 0.927

REP-5
{'021_bleach_cleanser': 0.25785417901600477, '006_mustard_bottle': 0.18202247191011237, '010_potted_meat_can': 0.11574344023323616}
ADD-10
{'021_bleach_cleanser': 0.8639596917605217, '006_mustard_bottle': 0.6831460674157304, '010_potted_meat_can': 0.5239067055393586}

[223d]

Hi, thanks for your remarkable work. I met the similar difficulty as @julyiii. The evaluation metrics didn't perform well. The object ADD seems ok, but the hand estimation is not quite well. Here's my logs at 50 epochs. I only changed the train_batch and workers to fit my hardware settings.

====== Options ======
HO3D_root: ../HandOccNet/data/HO3D/data
blocks: 1
channels: 256
epochs: 60
evaluate: False
host_folder: official
inp_res: 512
lambda_joints2d: 100.0
lambda_objects: 500.0
lr: 0.0001
lr_decay_gamma: 0.7
lr_decay_step: 10
mano_lambda_joints3d: 10000.0
mano_lambda_manopose: 10
mano_lambda_manoshape: 0.1
mano_lambda_regulpose: 1
mano_lambda_regulshape: 100.0
mano_lambda_verts3d: 10000.0
mano_neurons: [1024, 512]
mano_root: assets/mano_models
manual_seed: 0
momentum: 0.9
network: honet_transformer
obj_model_root: assets/object_models
resume: None
save_results: True
snapshot: 1
stacks: 1
test_batch: 16
test_freq: 10
train_batch: 8
transformer_depth: 1
transformer_head: 1
use_cuda: 1
weight_decay: 0.0005
workers: 12
=====================

Here's the results at 50 epochs.

Evaluation 3D KP results:
auc=0.334, mean_kp3d_avg=4.68 cm
Evaluation 3D KP PROCRUSTES ALIGNED results:
auc=0.767, mean_kp3d_avg=1.17 cm
Evaluation 3D KP SCALE-TRANSLATION ALIGNED results:
auc=0.344, mean_kp3d_avg=4.66 cm

Evaluation 3D MESH results:
auc=0.344, mean_kp3d_avg=4.52 cm
Evaluation 3D MESH ALIGNED results:
auc=0.778, mean_kp3d_avg=1.11 cm

F-scores
F@5.0mm = 0.164   F_aligned@5.0mm = 0.466
F@15.0mm = 0.534  F_aligned@15.0mm = 0.927

REP-5
{'021_bleach_cleanser': 0.25785417901600477, '006_mustard_bottle': 0.18202247191011237, '010_potted_meat_can': 0.11574344023323616}
ADD-10
{'021_bleach_cleanser': 0.8639596917605217, '006_mustard_bottle': 0.6831460674157304, '010_potted_meat_can': 0.5239067055393586}

May I ask how you solved this problem?@karta2155802

[karta2155802]

I'm not sure how I solved this problem, because it has been a year. In my memory, the training outcome is not quite stable. After training multiple times, I got similar results to the paper. Maybe you can try the latest paper HFL-Net. It has a similar structure to Semi-HO. But there are some errors in their code that need to be fixed by yourself.

[223d]

I'm not sure how I solved this problem, because it has been a year. In my memory, the training outcome is not quite stable. After training multiple times, I got similar results to the paper. Maybe you can try the latest paper HFL-Net. It has a similar structure to Semi-HO. But there are some errors in their code that need to be fixed by yourself.

Thank You for your reply!

[223d]

Okk, I will try to train with more epoches to find if the result will be better.Thank you for your reply! Hi,May I ask how many epochs did you train to achieve the results in your paper? @julyiii