Tiny-Sam vs HQ-Sam

[Meteplus]

with four points, the tiny-sam does'nt segment correctly,while HQ-Sam works well even for pictures with great noises

[xinghaochen]

Could you provide the original input image so that we could perform further analysis? Thanks.

[Meteplus]

give me your email address. I can send you my test images

[xinghaochen]

Thanks for your test images.

• I notice that the areas around the given prompt points are quite similar to that of background, i.e., both are in white, which make the segmentation of such mask quite difficult. Segmenting the main area of the image by TinySAM obtains reasonable results:
• You can also provide additional negative prompt points to obtain the following result:

  input_point = np.array([[1300, 100], [200, 300], [2400, 300], [1300, 1100], [100, 100], [2500, 1200]])
  input_label = np.array([1, 1, 1, 1, 0, 0])

  

• Your above results of HQ-SAM are obtained with a relative large model, i.e., ViT-L backbone, which is much larger than TinySAM. The Light HQ-SAM obtains the following result:

[Meteplus]

yes,i've got better results with negative points,thank you for your reply