small object is not good?

[CPFelix]

Thanks for the great work!
When I use the pretrained model(like ES-RV-L ES-TV-L ES-TV-M-MC-S1),I found that when objects in the image are small,then the model seg nothing.
Is it the loss caused by distill and is normal? @SimonZeng7108
test image:
--prompt "car"

[colinlin1982]

问题与 #17 相似。

目前可以生成稳定结果的蒸馏模型是litetext分支下的efficient_sam3_image_encoder_mobileclip_s1_ctx16，我用你这张图像（尺寸压缩到1/4）试了一下：

[SimonZeng7108]

Yes, unfortunately the efficient sam3 image encoders still underperforming at stage 1. For your problem, maybe it's worth finetuning on this dataset - Tracking-Any-Granularity.

[CPFelix]

Thanks for your reply! @colinlin1982
I try your advice and the cars are seg out. But I still want seg "water" for this image, and I also found low conf question and wrong results(I change conf_threhold=0.1):
boxes 置信度:[0.1080, 0.1108, 0.1586, 0.1013]

I think the problem is text_encoder lost some information about rare category(No coco category).
The official sam3(conf_threshold=0.4):
masks 数量: 2

• boxes 数量: 2
• boxes 置信度: tensor([0.5713, 0.4380]
  

[colinlin1982]

@CPFelix 我有个建议，只针对你这个应用场景，你可以试试提示词"muddy water"。

[CPFelix]

谢谢你的建议 @colinlin1982 我试了试确实能检测出了。看来类别的描述也很重要。