Doo you have the trained pointnet2_sem_seg model ?

[kkyan1995]

@charlesq34 , Do you have the trained pointnet2_sem_seg model ? can you give me it? my email is juanyan_sues@163.com , thank you very much!

[TianyangChen357]

I was training it. I can try to provide to you if still needed

[Leerw]

I have encountered a problem after I trained a model, the miou is fine, but when I restore the model and run test.py, I got nothing but the model segment the whole point cloud into the same class. Can you help me?

[TianyangChen357]

Hey there，

I will contact you later tonight since I m working on my final now.

Best， Tianyang

leerw notifications@github.com于2019年4月30日 周二上午5:11写道：

I have encountered a problem after I trained a model, the miou is fine, but when I restore the model and run test.py, I got nothing but the model segment the whole point cloud into the same class. Can you help me?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/charlesq34/pointnet2/issues/97#issuecomment-487874142, or mute the thread https://github.com/notifications/unsubscribe-auth/ALIBLWZO4YTWWPYXOYTXPNDPTAEK3ANCNFSM4GL5MCUA .

[BenChenCh]

@cagis2019 which dataset do you use to train pointnet2_sem_seg? could you prepare the dataset h5 file by your self or using the provided one?

[KevinYuk]

Hi @cagis2019,

Have you successfully trained the pointnet2_sem_seg model? Could you please share it with us? My email is SeedKunY@gmail.com

Thank you very much!

[TianyangChen357]

Hey there!

I did not try that yet but you can try this link, where they re-organized the way to utilize pointnet2. https://github.com/intel-isl/Open3D-PointNet2-Semantic3D (dataset and results refer to http://www.semantic3d.net/view_method_detail.php?method=PointNet2_Demo)

If you have any questions pertaining to Pointnet 1, I can help out.

Keep in touch, Tianyang

On Thu, Nov 21, 2019 at 9:21 PM KevinYuk notifications@github.com wrote:

Hi @cagis2019 https://github.com/cagis2019,

Have you successfully trained the pointnet2_sem_seg model? Could you please share it with us? My email is SeedKunY@gmail.com SeedKunY@gmail.com

Thank you very much!

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/charlesq34/pointnet2/issues/97?email_source=notifications&email_token=ALIBLW2XWUCYM6PAEFKNZIDQU463LA5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4JYGI#issuecomment-557358105, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLWZY77G3EBWV62ITNDDQU463LANCNFSM4GL5MCUA .

[KevinYuk]

Hey there! I did not try that yet but you can try this link, where they re-organized the way to utilize pointnet2. https://github.com/intel-isl/Open3D-PointNet2-Semantic3D (dataset and results refer to http://www.semantic3d.net/view_method_detail.php?method=PointNet2_Demo) If you have any questions pertaining to Pointnet 1, I can help out. Keep in touch, Tianyang … On Thu, Nov 21, 2019 at 9:21 PM KevinYuk @.> wrote: Hi @cagis2019 https://github.com/cagis2019, Have you successfully trained the pointnet2_sem_seg model? Could you please share it with us? My email is @. @.**> Thank you very much! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#97?email_source=notifications&email_token=ALIBLW2XWUCYM6PAEFKNZIDQU463LA5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4JYGI#issuecomment-557358105>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLWZY77G3EBWV62ITNDDQU463LANCNFSM4GL5MCUA .

Hi @cagis2019,

Thanks for your nice reply.

PointNet1 pre-trained model is OK. We are trying to design an experiment by using pointnet1 (or pointnet2). If my understanding is right, you mean you can share the pre-trained model of pointnet1 with us? right?

Thanks again for your reply.

[TianyangChen357]

Are you trying to use transfer learning, as you mentioned pre-trained model? I am currently also working on this. Yes, I can share but that may not work very well, which is assumed overfitting (90% OA for training but ~30% OA for testing). I trained the model on our manually labeled dataset (LiDAR data or outdoor scenes related to bridges).

I am currently trying to make a pre-trained model on the dataset I identified in the last email. Hopefully, I can leverage that benchmark data on my situation.

I can share that with you tomorrow since that is in the Lab if you still need it.

With my best regards, Tianyang

On Thu, Nov 21, 2019 at 9:49 PM KevinYuk notifications@github.com wrote:

Hey there! I did not try that yet but you can try this link, where they re-organized the way to utilize pointnet2. https://github.com/intel-isl/Open3D-PointNet2-Semantic3D (dataset and results refer to http://www.semantic3d.net/view_method_detail.php?method=PointNet2_Demo) If you have any questions pertaining to Pointnet 1, I can help out. Keep in touch, Tianyang … <#m-2487194901860533358> On Thu, Nov 21, 2019 at 9:21 PM KevinYuk @.> wrote: Hi @cagis2019 https://github.com/cagis2019 https://github.com/cagis2019 https://github.com/cagis2019, Have you successfully trained the pointnet2_sem_seg model? Could you please share it with us? My email is @. @.**> Thank you very much! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#97 https://github.com/charlesq34/pointnet2/issues/97?email_source=notifications&email_token=ALIBLW2XWUCYM6PAEFKNZIDQU463LA5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4JYGI#issuecomment-557358105>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLWZY77G3EBWV62ITNDDQU463LANCNFSM4GL5MCUA .

Hi @cagis2019 https://github.com/cagis2019,

Thanks for your nice reply.

PointNet1 pre-trained model is OK. We are trying to design an experiment by using pointnet1 (or pointnet2). If my understanding is right, you mean you can share the pre-trained model of pointnet1 with us? right?

Thanks again for your reply.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/charlesq34/pointnet2/issues/97?email_source=notifications&email_token=ALIBLW52RTN5GFG3PQAZFG3QU5CFDA5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4LILY#issuecomment-557364271, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLWZGGTWVLUH2UR3JDU3QU5CFDANCNFSM4GL5MCUA .

[KevinYuk]

@cagis2019 , Yes!!! We are trying to use transfer learning.

If you can share the pre-trained model with us, it would be great. We will try it based on kitti dataset: data_scene_flow which has been transferred to LiDAR point cloud data. It's a small outdoor dataset which only mainly include car object.

Actually, we have tried rangenet for this before, however, it doesn't work fine. Hope your pre-trained model works fine.

Thanks a lot for your help.

[TianyangChen357]

hey there, I have sent models to your email. You can check that

KevinYuk notifications@github.com于2019年11月21日 周四下午10:04写道：

@cagis2019 https://github.com/cagis2019 , Yes!!! We are trying to use transfer learning.

If you can share the pre-trained model with us, it would be great. We will try it based on kitti dataset: data_scene_flow which has been transferred to LiDAR point cloud data. It's a small outdoor dataset which only mainly include car object.

Actually, we have tried rangenet for this before, however, it doesn't work fine. Hope your pre-trained model works fine.

Thanks a lot for your help.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/charlesq34/pointnet2/issues/97?email_source=notifications&email_token=ALIBLWZGXXVON7GQIDY4N33QU5D35A5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4MB3A#issuecomment-557367532, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLW64FWPGT5AFG6HHE63QU5D35ANCNFSM4GL5MCUA .

[TianyangChen357]

Hi Kun,

Do you have any progress on transfer learning? I stacked there and did not know how to start the transfer learning with Pointnet.

Any instructions?

Best, Tianyang

On Fri, Nov 22, 2019 at 10:29 PM Tianyang Chen tchen19@uncc.edu wrote:

hey there, I have sent models to your email. You can check that

KevinYuk notifications@github.com于2019年11月21日 周四下午10:04写道：

@cagis2019 https://github.com/cagis2019 , Yes!!! We are trying to use transfer learning.

If you can share the pre-trained model with us, it would be great. We will try it based on kitti dataset: data_scene_flow which has been transferred to LiDAR point cloud data. It's a small outdoor dataset which only mainly include car object.

Actually, we have tried rangenet for this before, however, it doesn't work fine. Hope your pre-trained model works fine.

Thanks a lot for your help.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/charlesq34/pointnet2/issues/97?email_source=notifications&email_token=ALIBLWZGXXVON7GQIDY4N33QU5D35A5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4MB3A#issuecomment-557367532, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLW64FWPGT5AFG6HHE63QU5D35ANCNFSM4GL5MCUA .

[MartinMao1101]

Hi Kun, Do you have any progress on transfer learning? I stacked there and did not know how to start the transfer learning with Pointnet. Any instructions? Best, Tianyang … On Fri, Nov 22, 2019 at 10:29 PM Tianyang Chen @.> wrote: hey there, I have sent models to your email. You can check that KevinYuk @.>于2019年11月21日 周四下午10:04写道： > @cagis2019 https://github.com/cagis2019 , > Yes!!! We are trying to use transfer learning. > > If you can share the pre-trained model with us, it would be great. > We will try it based on kitti dataset: data_scene_flow which has been > transferred to LiDAR point cloud data. It's a small outdoor dataset which > only mainly include car object. > > Actually, we have tried rangenet for this before, however, it doesn't > work fine. Hope your pre-trained model works fine. > > Thanks a lot for your help. > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <#97?email_source=notifications&email_token=ALIBLWZGXXVON7GQIDY4N33QU5D35A5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4MB3A#issuecomment-557367532>, > or unsubscribe > https://github.com/notifications/unsubscribe-auth/ALIBLW64FWPGT5AFG6HHE63QU5D35ANCNFSM4GL5MCUA > . >

Hi Tianyang

i am using PN2 for my thesis and found your excellent job and the answers quite helpful. i have pre-trained the model with the given official dataset ShapeNet-Part and got a good result, then i ran the model again with my own dataset but the result was realy bad. Now i am supposed to use transfer learning to apply the pre-trained parameter for my own dataset. My problem is that i dont quite know how to do that, could you please share your instruction about how to use transfer learning?

With my best regards Ning

[TianyangChen357]

Transfer learning is to utilize a trained network as the initial network for your model, otherwise, your model will be trained from scratch (with a random initial network). Such a trained network should better be trained on a big dataset (e.g., benchmark dataset. PS: I used semantic-8 dataset, http://www.semantic3d.net/), which should better be similar to your domain (see more explanation of transfer learning https://machinelearningmastery.com/transfer-learning-for-deep-learning/). For example, my model is focused on hydraulic structures in rural areas; thus, I used the semantic-8 dataset to train a network as the initial network for my model, which is also for outdoor scenes that may help to detect ground and vegetation. Technically, you will load the trained network as the initial network before training. Then lock the weights for some of the early layers (near to the input layer) and unlock the weights for some latter layers (near to the output layer). How many layers to unlock? You may have to have a sensitivity analysis to know about it but you can start with 1 - 10.

With many best wishes, Tianyang

On Mon, Oct 25, 2021 at 8:54 AM MartinMao1101 @.***> wrote:

[Caution: Email from External Sender. Do not click or open links or attachments unless you know this sender.]

Hi Kun, Do you have any progress on transfer learning? I stacked there and did not know how to start the transfer learning with Pointnet. Any instructions? Best, Tianyang … <#m_-3498671832641642997_m890185186136466148> On Fri, Nov 22, 2019 at 10:29 PM Tianyang Chen @.> wrote: hey there, I have sent models to your email. You can check that KevinYuk @.>于2019年11月21日 周四下午10:04写道： > @cagis2019 https://github.com/cagis2019 , > Yes!!! We are trying to use transfer learning. > > If you can share the pre-trained model with us, it would be great. > We will try it based on kitti dataset: data_scene_flow which has been > transferred to LiDAR point cloud data. It's a small outdoor dataset which > only mainly include car object. > > Actually, we have tried rangenet for this before, however, it doesn't > work fine. Hope your pre-trained model works fine. > > Thanks a lot for your help. > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <#97 https://github.com/charlesq34/pointnet2/issues/97?email_source=notifications&email_token=ALIBLWZGXXVON7GQIDY4N33QU5D35A5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4MB3A#issuecomment-557367532>,

or unsubscribe > https://github.com/notifications/unsubscribe-auth/ALIBLW64FWPGT5AFG6HHE63QU5D35ANCNFSM4GL5MCUA . >

Hi Tianyang

i am using PN2 for my thesis and found your excellent job and the answers quite helpful. i have pre-trained the model with the given official dataset ShapeNet-Part and got a good result, then i ran the model again with my own dataset but the result was realy bad. Now i am supposed to use transfer learning to apply the pre-trained parameter for my own dataset. My problem is that i dont quite know how to do that, could you please share your instruction about how to use transfer learning?

With my best regards Ning

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/charlesq34/pointnet2/issues/97#issuecomment-950894188, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLW6EWLYFYDMOOHV5DTLUIVHOTANCNFSM4GL5MCUA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

[MartinMao1101]

Transfer learning is to utilize a trained network as the initial network for your model, otherwise, your model will be trained from scratch (with a random initial network). Such a trained network should better be trained on a big dataset (e.g., benchmark dataset. PS: I used semantic-8 dataset, http://www.semantic3d.net/), which should better be similar to your domain (see more explanation of transfer learning https://machinelearningmastery.com/transfer-learning-for-deep-learning/). For example, my model is focused on hydraulic structures in rural areas; thus, I used the semantic-8 dataset to train a network as the initial network for my model, which is also for outdoor scenes that may help to detect ground and vegetation. Technically, you will load the trained network as the initial network before training. Then lock the weights for some of the early layers (near to the input layer) and unlock the weights for some latter layers (near to the output layer). How many layers to unlock? You may have to have a sensitivity analysis to know about it but you can start with 1 - 10. With many best wishes, Tianyang On Mon, Oct 25, 2021 at 8:54 AM MartinMao1101 @.**> wrote: … [Caution: Email from External Sender. Do not click or open links or attachments unless you know this sender.] Hi Kun, Do you have any progress on transfer learning? I stacked there and did not know how to start the transfer learning with Pointnet. Any instructions? Best, Tianyang … <#m_-3498671832641642997_m890185186136466148> On Fri, Nov 22, 2019 at 10:29 PM Tianyang Chen @.> wrote: hey there, I have sent models to your email. You can check that KevinYuk @.>于2019年11月21日 周四下午10:04写道： > @cagis2019 https://github.com/cagis2019 , > Yes!!! We are trying to use transfer learning. > > If you can share the pre-trained model with us, it would be great. > We will try it based on kitti dataset: data_scene_flow which has been > transferred to LiDAR point cloud data. It's a small outdoor dataset which > only mainly include car object. > > Actually, we have tried rangenet for this before, however, it doesn't > work fine. Hope your pre-trained model works fine. > > Thanks a lot for your help. > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <#97 <#97>?email_source=notifications&email_token=ALIBLWZGXXVON7GQIDY4N33QU5D35A5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4MB3A#issuecomment-557367532>, > or unsubscribe > https://github.com/notifications/unsubscribe-auth/ALIBLW64FWPGT5AFG6HHE63QU5D35ANCNFSM4GL5MCUA > . > Hi Tianyang i am using PN2 for my thesis and found your excellent job and the answers quite helpful. i have pre-trained the model with the given official dataset ShapeNet-Part and got a good result, then i ran the model again with my own dataset but the result was realy bad. Now i am supposed to use transfer learning* to apply the pre-trained parameter for my own dataset. My problem is that i dont quite know how to do that, could you please share your instruction about how to use transfer learning? With my best regards Ning — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#97 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLW6EWLYFYDMOOHV5DTLUIVHOTANCNFSM4GL5MCUA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

Hello Tianyang, thank you so much for the detailed explanation! These days i have pre-trained my model with a bigger dataset and get a better result, thanks to your instruction! After the pre-training, i found the location of pre-trained model, correspondingly there is another "train.py". Actually my problem is how can i set which layers to be locked and which are set free? Thank you very much again! Best, Ning

[TianyangChen357]

I cannot help more on the operation level in Tensorflow. I used the other network (ConvPoint) eventually for my study, which is developed on PyTorch. I did transfer learning there. I did not have a try at using TensorFlow but I am sure you can find a lot of tutorials for that. An official tutorial is attached here for your reference: https://www.tensorflow.org/tutorials/images/transfer_learning

With many best wishes, Tianyang

On Wed, Nov 3, 2021 at 4:46 AM MartinMao1101 @.***> wrote:

[Caution: Email from External Sender. Do not click or open links or attachments unless you know this sender.]

Transfer learning is to utilize a trained network as the initial network for your model, otherwise, your model will be trained from scratch (with a random initial network). Such a trained network should better be trained on a big dataset (e.g., benchmark dataset. PS: I used semantic-8 dataset, http://www.semantic3d.net/), which should better be similar to your domain (see more explanation of transfer learning https://machinelearningmastery.com/transfer-learning-for-deep-learning/). For example, my model is focused on hydraulic structures in rural areas; thus, I used the semantic-8 dataset to train a network as the initial network for my model, which is also for outdoor scenes that may help to detect ground and vegetation. Technically, you will load the trained network as the initial network before training. Then lock the weights for some of the early layers (near to the input layer) and unlock the weights for some latter layers (near to the output layer). How many layers to unlock? You may have to have a sensitivity analysis to know about it but you can start with 1 - 10. With many best wishes, Tianyang On Mon, Oct 25, 2021 at 8:54 AM MartinMao1101 @.*

> wrote: … <#m-805538229071222687> [Caution: Email from External Sender. Do not click or open links or attachments unless you know this sender.] Hi Kun, Do you have any progress on transfer learning? I stacked there and did not know how to start the transfer learning with Pointnet. Any instructions? Best, Tianyang … <#m_-3498671832641642997_m890185186136466148> On Fri, Nov 22, 2019 at 10:29 PM Tianyang Chen @.> wrote: hey there, I have sent models to your email. You can check that KevinYuk @.>于2019年11月21日 周四下午10:04写道： > @cagis2019 https://github.com/cagis2019 , > Yes!!! We are trying to use transfer learning. > > If you can share the pre-trained model with us, it would be great. > We will try it based on kitti dataset: data_scene_flow which has been > transferred to LiDAR point cloud data. It's a small outdoor dataset which > only mainly include car object. > > Actually, we have tried rangenet for this before, however, it doesn't > work fine. Hope your pre-trained model works fine. > > Thanks a lot for your help. > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <#97 https://github.com/charlesq34/pointnet2/issues/97 <#97 https://github.com/charlesq34/pointnet2/issues/97>?email_source=notifications&email_token=ALIBLWZGXXVON7GQIDY4N33QU5D35A5CNFSM4GL5MCUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4MB3A#issuecomment-557367532>,

or unsubscribe > https://github.com/notifications/unsubscribe-auth/ALIBLW64FWPGT5AFG6HHE63QU5D35ANCNFSM4GL5MCUA . > Hi Tianyang i am using PN2 for my thesis and found your excellent job and the answers quite helpful. i have pre-trained the model with the given official dataset ShapeNet-Part and got a good result, then i ran the model again with my own dataset but the result was realy bad. Now i am supposed to use transfer learning to apply the pre-trained parameter for my own dataset. My problem is that i dont quite know how to do that, could you please share your instruction about how to use transfer learning? With my best regards Ning — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#97 (comment) https://github.com/charlesq34/pointnet2/issues/97#issuecomment-950894188>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLW6EWLYFYDMOOHV5DTLUIVHOTANCNFSM4GL5MCUA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub .

Hello Tianyang, thank you so much for the detailed explanation! These days i have pre-trained my model with a bigger dataset and get a better result, thanks to your instruction! After the pre-training, i found the location of pre-trained model, correspondingly there is another "train.py". Actually my problem is how can i set which layers to be locked and which are set free? Thank you very much again! Best, Ning

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/charlesq34/pointnet2/issues/97#issuecomment-958749267, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALIBLW2KMJQCXP4VLAPO2NDUKDZF3ANCNFSM4GL5MCUA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.