dataset for segmenting marine animals be placed?

[Lei0829]

Hello, how should the dataset for segmenting marine animals be placed?

[xiongxyowo]

Hi, following the existing settings, MAS3K and RMAS are not trained together. In other words, we train one model on the training set of MAS3K and tested it on the test set of MAS3K, while also training another model on RMAS.

[Lei0829]

Hi, following the existing settings, MAS3K and RMAS are not trained together. In other words, we train one model on the training set of MAS3K and tested it on the test set of MAS3K, while also training another model on RMAS.

我使用MAS3K的数据集，测试出来的效果和原先图片标注的效果差距很大

[xiongxyowo]

Hi, since the network is trained at 352 resolution, it is normal for predictions to look blurrier than GT in some cases. You can also check our pre-computed maps in google drive.

[Lei0829]

Hi, since the network is trained at 352 resolution, it is normal for predictions to look blurrier than GT in some cases. You can also check our pre-computed maps in google drive.

会不会是我训练轮数按你给我的20次过少，改成200--epoch 200
--lr 0.001
--batch_size 12
还需要改动什么么

[xiongxyowo]

Hi, SAM2-UNet generally converges within 20 epochs. Thus, increasing the number of epochs might not be helpful. Although the example you present seems not good, SAM2-UNet still attains state-of-the-art performance in the task of marine animal segmentation. You can design more advanced segmentation decoders to obtain better segmentation results.

[Lei0829]

Hi, SAM2-UNet generally converges within 20 epochs. Thus, increasing the number of epochs might not be helpful. Although the example you present seems not good, SAM2-UNet still attains state-of-the-art performance in the task of marine animal segmentation. You can design more advanced segmentation decoders to obtain better segmentation results.

当我在尝试使用200epoch时候，两个指标提高了两个点，这是过拟合了？