training and weights

[imu-jan]

After I reproduced, the results split on davis2017 were 10 percentage points lower than the results you gave. Why is this? Can you give it weight after completing the training?

[qinliuliuqin]

@imu-jan Hi, can you provide more context on your reproduction? Which table were you referring to? Were you retraining the code or simply using my weights? I will update the links of my weights this weekend.

[imu-jan]

嗨，您能提供更多关于您的复制的背景信息吗？你指的是哪张桌子？你是在重新训练代码还是只是使用我的权重？我将在本周末更新我的重量链接。

I reproduce on a 4090gpu with the pre-training batch_size set to 16 and the main training batch_size set to 4. The j&f on davis2017 using trained weights is 73.4, which is much lower than the 84.4 in the paper

[imu-jan]

livos supports 4096p high-resolution segmentation, how should I do it? I don't see a way to give 4096p inference in your code. I want to reason about the davis2017 dataset in 4096p resolution, should I prepare the 4096p resolution dataset or can I just enter the 480p dataset in your code and resize it?

[imu-jan]

livos supports 4096p high-resolution segmentation, how should I do it? I don't see a way to give 4096p inference in your code. I want to reason about the davis2017 dataset in 4096p resolution, should I prepare the 4096p resolution dataset or can I just enter the 480p dataset in your code and resize it?

[qinliuliuqin]

嗨，您能提供更多关于您的复制的背景信息吗？你指的是哪张桌子？你是在重新训练代码还是只是使用我的权重？我将在本周末更新我的重量链接。

I reproduce on a 4090gpu with the pre-training batch_size set to 16 and the main training batch_size set to 4. The j&f on davis2017 using trained weights is 73.4, which is much lower than the 84.4 in the paper

Hi, I don't think you followed our training recipe. As mentioned in Sec. 4.1, we omitted pretraining on static images and BL30K. For the main training, we set the batch size to 16 with 125K training iterations. You should be able to reproduce the results if you follow our code and training recipe.

[qinliuliuqin]

livos supports 4096p high-resolution segmentation, how should I do it? I don't see a way to give 4096p inference in your code. I want to reason about the davis2017 dataset in 4096p resolution, should I prepare the 4096p resolution dataset or can I just enter the 480p dataset in your code and resize it?

For our experiments, we simply upsample DAVIS videos to 4096p for training and evaluation. To reproduce the results, you may need to train it longer and also gradually finetune a model trained on a coarser resolution. For example, to train a 4096p model, it's better to finetune it on a 3072p model rather than a 480p model.

[qinliuliuqin]

@imu-jan Hi, links for model weights have been updated. Let me know if you have any other questions.