Zigeng commited on
Commit
d67c795
1 Parent(s): 819dc58

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +2 -33
README.md CHANGED
@@ -6,15 +6,13 @@ tags:
6
  ---
7
 
8
  # SlimSAM: 0.1% Data Makes Segment Anything Slim
9
- <div align="center">
10
- <img src="images/paper/intro.PNG" width="66%">
11
- <img src="images/paper/everything.PNG" width="100%">
12
- </div>
13
 
14
  > **0.1% Data Makes Segment Anything Slim**
15
  > [Zigeng Chen](https://github.com/czg1225), [Gongfan Fang](https://fangggf.github.io/), [Xinyin Ma](https://horseee.github.io/), [Xinchao Wang](https://sites.google.com/site/sitexinchaowang/)
16
  > [Learning and Vision Lab](http://lv-nus.org/), National University of Singapore
17
  > Paper: [[Arxiv]](https://arxiv.org/abs/2312.05284)
 
18
 
19
  ## Introduction
20
 
@@ -24,38 +22,9 @@ tags:
24
 
25
  **SlimSAM** is a novel SAM compression method, which efficiently reuses pre-trained SAMs without the necessity for extensive retraining. This is achieved by the efficient reuse of pre-trained SAMs through a unified pruning-distillation framework. To enhance knowledge inheritance from the original SAM, we employ an innovative alternate slimming strategy that partitions the compression process into a progressive procedure. Diverging from prior pruning techniques, we meticulously prune and distill decoupled model structures in an alternating fashion. Furthermore, a novel label-free pruning criterion is also proposed to align the pruning objective with the optimization target, thereby boosting the post-distillation after pruning.
26
 
27
- ![Frame](images/paper/frame.PNG?raw=true)
28
-
29
  SlimSAM achieves approaching performance while reducing the parameter counts to **0.9\% (5.7M)**, MACs to **0.8\% (21G)**, and requiring mere **0.1\% (10k)** of the training data when compared to the original SAM-H. Extensive experiments demonstrate that our method realize significant superior performance while utilizing over **10 times** less training data when compared to other SAM compression methods.
30
 
31
- ## Visualization Results
32
-
33
- Qualitative comparison of results obtained using point prompts, box prompts, and segment everything prompts are shown in the following section.
34
 
35
- ### Segment Everything Prompts
36
- <div align="center">
37
- <img src="images/paper/everything2.PNG" width="100%">
38
- </div>
39
-
40
- ### Box Prompts and Point Prompts
41
- <div align="center">
42
- <img src="images/paper/prompt.PNG" width="100%">
43
- </div>
44
-
45
-
46
- ## Quantitative Results
47
-
48
- We conducted a comprehensive comparison encompassing performance, efficiency, and training costs with other SAM compression methods and structural pruning methods.
49
-
50
- ### Comparing with other SAM compression methods.
51
- <div align="center">
52
- <img src="images/paper/compare_tab1.PNG" width="100%">
53
- </div>
54
-
55
- ### Comparing with other structural pruning methods.
56
- <div align="center">
57
- <img src="images/paper/compare_tab2.PNG" width="50%">
58
- </div>
59
 
60
 
61
 
 
6
  ---
7
 
8
  # SlimSAM: 0.1% Data Makes Segment Anything Slim
9
+
 
 
 
10
 
11
  > **0.1% Data Makes Segment Anything Slim**
12
  > [Zigeng Chen](https://github.com/czg1225), [Gongfan Fang](https://fangggf.github.io/), [Xinyin Ma](https://horseee.github.io/), [Xinchao Wang](https://sites.google.com/site/sitexinchaowang/)
13
  > [Learning and Vision Lab](http://lv-nus.org/), National University of Singapore
14
  > Paper: [[Arxiv]](https://arxiv.org/abs/2312.05284)
15
+ > Code: [[SlimSAM]](https://github.com/czg1225/SlimSAM)
16
 
17
  ## Introduction
18
 
 
22
 
23
  **SlimSAM** is a novel SAM compression method, which efficiently reuses pre-trained SAMs without the necessity for extensive retraining. This is achieved by the efficient reuse of pre-trained SAMs through a unified pruning-distillation framework. To enhance knowledge inheritance from the original SAM, we employ an innovative alternate slimming strategy that partitions the compression process into a progressive procedure. Diverging from prior pruning techniques, we meticulously prune and distill decoupled model structures in an alternating fashion. Furthermore, a novel label-free pruning criterion is also proposed to align the pruning objective with the optimization target, thereby boosting the post-distillation after pruning.
24
 
 
 
25
  SlimSAM achieves approaching performance while reducing the parameter counts to **0.9\% (5.7M)**, MACs to **0.8\% (21G)**, and requiring mere **0.1\% (10k)** of the training data when compared to the original SAM-H. Extensive experiments demonstrate that our method realize significant superior performance while utilizing over **10 times** less training data when compared to other SAM compression methods.
26
 
 
 
 
27
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28
 
29
 
30