forked from SysCV/sam-hq
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sam-hq.txt
102 lines (68 loc) · 3.25 KB
/
sam-hq.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
# Would be really good to know about the donuts
2) Can I improve the results alot by pinpointing several points, perhaps?
Not really.
Not sure how to select all the donuts to be honest. Seems hard for it
# Read the summary paper of work done on Sam
# Note SAM is more sensitive to textures than shapes!
Which is why it's poor at finding shapes!
# Sam in High Quality
https://arxiv.org/pdf/2306.01567.pdf
https://github.com/SysCV/SAM-HQ
# Lama
LaMa: Resolution-robust Large Mask Inpainting with Fourier Convolutions
https://github.com/advimman/lama
# Inpaint Anything
https://arxiv.org/abs/2304.06790
https://github.com/geekyutao/Inpaint-Anything
> Remove Anything
> Fill Anything
> Replace Anything
# Restore Anything (No code available yet on the github!)
https://arxiv.org/pdf/2305.13093.pdf
https://github.com/eth-siplab/RAP
# Deshadow-Anything
https://arxiv.org/pdf/2309.11715
(No code available, very short paper)
# Track Anything (Video SAM)
Segment Anything Meets Video
https://arxiv.org/abs/2304.11968
https://github.com/gaomingqi/Track-Anything
# AI Summary
A curated list of general AI methods for Anything: AnyObject, AnyGeneration, AnyModel, AnyTask, etc.
https://github.com/VainF/Awesome-Anything
# One big issues is that the CUDA memory is never released, so can only run a few times before graphics card is out of memory
export PYTORCH_CUDA_ALLOC_CONF="max_split_size_mb=10"
# Output files are just written to demo/hq_sam_result + demo/hq_sam_tiny_result
So, to view my results, just need to add my images into the system, but also need to set specific places for the input masks.......
# Stuff to do:
1. https://github.com/SysCV/sam-hq/tree/main - go through all the stuff here
2. Go through the issues page:
https://github.com/SysCV/sam-hq/issues?q=is%3Aopen+is%3Aissue
(And the closed issues too)
Basically, I want to see visually how it does on my images, that's all, so what's the easiest way to see it
Also, how fast it does embedding for my images on my local machine, not in cloud
# Then next question is how easy to just pop old model out and pop new model into my code
How compatible are the models, methods?
# Why is there a visual demo directory in the git? Are these the source images?
No. These are the before and after animated gifs of SAM vs SAM-HQ
# hq_token_only = False / True
To achieve best visualization effect, for images contain multiple objects (like typical coco images), we suggest to set hq_token_only=False
For images contain single object, we suggest to set hq_token_only = True
# Currently need to be all .png files or all .jpg files
????
Stick with .jpg for now? Or will only .png files convert?
# SO I need 8 jpeg examples
Can convert from jpeg obviously
Then do points or bounding rectangles
Can do in art package quickly enough
Which art package? Gwenview is ok
Just pick 8 examples first:
0) bar-lady.jpg - start (155,32) size (414,768)
1) bracelet.jpg - start (45,99) size (869,681)
2) drink.jpg - one point at (167,285)
3) flower-lady.jpg one point at (310,130)
4) old-lady-robot.jpg one point at (62,112)
5) hi-res-flowers.jpg one point at (1130,1567)
6) johnny-haynes.jpg start (202,0) size (786,922)
7) donuts.jpg - try 2 points at (180,220) and (170,379)
Seems to be better than SAM, but not massively. Lots of things still fuck up