CUDA Segmentation Stats

Collect various statistics from a semantic segmentation datasets. The program will recursively search for label images and accumulate statistics. This program uses cnmem for allocation, to grab 95% of the available GPU memory on startup.

Notice: Always use the number of classes plus one. All values higher than number of classes (e.g. 255 for ignore) are grouped into the top class.

Performance: Using a TitanXp and 8 computing threads/streams I get around 36.8403 im/sec with resizing and 63.7906 im/sec without.

There are 3 available stats currently implemented:

Frequency

This is the total count for each label class.

example output:

class	count
0	5.46547e+07
1	1.74758e+07
2	4.44468e+08
3	461764
4	1.70721e+07
5	1.72539e+06
6	1.69774e+06
7	2.07309e+06
8	2.16049e+08
9	2.01027e+08
10	3.62953e+08
11	8.37151e+07
12	342664
13	2.44656e+07
14	3.25476e+06
15	1.51454e+06
16	366913
17	1.24194e+06
18	1.54103e+06
19	2.31194e+09
20	1.9688e+09

Image Frequency

This is the total count for each label for each image.

example output:

path	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20
img1.png	709962	351	806772	574	325	310	439	412	174552	510	214918	120	576	97949	51	44	81	72	62	9689	79383
img2.png	0	0	154973	126	135	192	331	141	14764	119	85	90	96	327	161	68	114	133	120	1.27981e+06	645368

Pixel Frequency

This is the frequency of a class for each pixel location. Note the output is in exr format. It can be easily read with opencv using the flag IMREAD_ANYDEPTH | IMREAD_ANYCOLOR.

Example output (Cityscapes sidewalk calss):

Build

cd /path/to/project
mkdir build
cmake ..
make

Usage

• -i: Input folder to recursively search under
  • E.g. -i /path/to/data
• -o: Output folder to store stats
  • E.g. -o stat_out/
• -e: Ending to filter by. No regex or wild cards supported. [Defalult: png]
  • E.g. -e labelIds.png
• -n: Number of threads to use. Each thread takes up space on GPU. [Default: 8]
  • E.g. -n 3
• -g: List of device numbers to use, comma separated. Currently only one gpu at a time is supported [Default: 0]
  • E.g. -g 0
• -c: Number of classes (Remember to add one class for ignore/255) [Default: 20]
  • E.g. -c 5
• -h: Integer height to resize to. Negative number means no resize. [Default: -1]
  • E.g. -h 1024
• -w: Integer height to resize to. Negative number means no resize. [Default: -1]
  • E.g. -w 1024
• -f: Compute Frequency statistic [Default: disabled]
  • E.g. -f
• -p: Compute Pixel Frequency statistic [Default: disabled]
  • E.g. -p
• -m: Compute Image Frequency statistic [Default: disabled]
  • E.g. -m
• -a: Compute all available statistics [Default: disabled]
  • E.g. -a

Adding all examples to a single command:

./SegStats -i /path/to/data -o stat_out/ -e labelIds.png -n 3 -g 0 -c 5 -h 1024 -w 1024 -f -p -m -a

Limitations

• All images must be the same size (or resized to be the size).
• Only one GPU is implemented.