TF-CCTV is a Coral compatible Tensorflow model designed to process CCTV footage.

118 commits 7 branches 7 tags 1.2 GiB

PureBasic 96.3%
Python 3.2%
TypeScript 0.3%

Find a file

Curid b2726f4a97 export yolov9s cpu		2025年01月31日 20:50:55 +00:00
bench	add yolo	2025年01月25日 17:17:46 +00:00
datasets	small fixes	2024年12月18日 20:33:23 +00:00
docker	add eval runner script	2025年01月27日 20:43:43 +00:00
docs	add accuracy plots	2025年01月31日 17:24:13 +00:00
eval	add missing evals	2025年01月29日 18:21:14 +00:00
labels	evaluate spaghettinet	2024年09月21日 18:51:09 +00:00
models	Fix cctv4 export_tflite.py	2025年01月27日 08:19:22 +00:00
od	Fix cctv4 export_tflite.py	2025年01月27日 08:19:22 +00:00
pretrained_models	evaluate most of the official tflite models	2024年11月29日 18:57:17 +00:00
py_utils	add accuracy plots	2025年01月31日 17:24:13 +00:00
records	add objects365 eval records to the repo	2024年12月02日 20:03:08 +00:00
utils	release cctv3.3	2023年01月04日 20:05:09 +00:00
yolo	export yolov9s cpu	2025年01月31日 20:50:55 +00:00
.flake8	add Fiftyone script	2021年12月06日 19:32:54 +01:00
.gitignore	remove USE_EDGETPU	2025年01月26日 18:54:28 +00:00
.pylintrc	add Fiftyone script	2021年12月06日 19:32:54 +01:00
LICENSE	change license to agplv3	2025年01月25日 17:07:34 +00:00
README.md	add accuracy plots	2025年01月31日 17:24:13 +00:00

README.md

TF-CCTV

Custom Coral compatible Tensorflow models designed to process CCTV footage.

Matrix chat

Accuracy

objects365-val-cctv-v1_person

EDGETPU

objects365-val-cctv-v1_person_edgetpu_color objects365-val-cctv-v1_person_edgetpu_gray

raw data

MODEL	mAP	L	M	S	RPI5
mobilenetV1	color	33.5	50.5	12.4	0.4	3.3	a
mobilenetV1	gray	29.6	45.8	9.1	0.2	3.3
-
mobilenetV2	color	38.2	57.9	13.8	0.9	3.7	b
mobilenetV2	gray	32.7	51.4	9.4	0.7	3.7
-
mobiledet_official	color	46.4	64.0	24.7	4.5	6.1	c
mobiledet_official	gray	41.4	59.3	18.3	3.0	6.1
-
mobiledet	color	45.3	62.7	23.6	4.4	4.0	d
mobiledet	gray	40.5	58.2	17.4	2.9	4.0	a
-
spaghettinet_m	color	45.5	62.5	25.0	3.3	4.1	e
spaghettinet_m	gray	40.6	57.5	19.7	2.8	4.1
spaghettinet_l	color	46.9	63.8	26.5	5.2	4.1	f
spaghettinet_l	gray	41.7	58.0	21.7	2.8	4.1	b
spaghettinet_l_official	color	45.9	63.4	24.4	3.8	3.7	g
spaghettinet_l_official	gray	41.1	58.2	19.9	2.0	3.7
-
yolov5nu_320x320	color	44.6	62.4	22.8	0.7	8.0	h
yolov5nu_320x320	gray	33.9	50.1	15.2	1.3	8.0	c
yolov5nu_480x480	color	51.9	68.3	31.3	7.3	19.0	i
yolov5nu_480x480	gray	38.8	54.5	21.5	3.3	19.0	d
yolov5nu_576x576	color	54.2	69.6	36.6	11.0	31.0	j
yolov5nu_576x576	gray	42.7	58.0	26.7	4.6	31.0
-
yolov5n6u_320x320	color	42.0	59.7	20.2	1.0	8.7	k
yolov5n6u_320x320	gray	31.4	46.7	14.2	0.7	8.7	e
-
yolov5su_320x320	color	50.6	67.0	30.9	5.5	10.2	h
yolov5su_320x320	gray	39.4	55.6	21.7	2.6	10.2	c
yolov5su_480x480	color	58.4	73.0	41.9	14.7	37.0	i
yolov5su_480x480	gray	47.1	62.5	31.0	8.2	37.0
-
yolov5s6u_320x320	color	47.8	65.8	25.8	3.0	13.2	k
yolov5s6u_320x320	gray	36.8	53.6	17.1	1.4	13.2	e
-
yolov5mu_320x320	color	59.5	75.3	41.1	10.4	21.5	h
yolov5mu_320x320	gray	51.4	69.1	29.9	4.9	21.5	c
-
yolov5m6u_320x320	color	55.6	73.0	34.7	5.1	27.7	k
yolov5m6u_320x320	gray	49.1	66.7	28.1	3.8	27.7
-
yolov5lu_320x320	color	61.2	76.4	44.3	11.6	33.0	h
yolov5lu_320x320	gray	54.3	71.0	35.0	7.8	33.0
-
yolov5l6u_320x320	color	57.2	74.9	36.4	6.3	45.3	k
yolov5l6u_320x320	gray	50.1	68.3	28.7	5.3	45.3
-
yolov8n_320x320	color	43.8	60.5	23.4	2.1	8.2	l
yolov8n_320x320	gray	29.4	43.9	13.4	1.1	8.2	f
yolov8n_480x480	color	50.3	65.6	32.3	8.5	19.1	m
yolov8n_480x480	gray	36.4	51.3	19.6	4.3	19.1	g
yolov8n_576x576	color	52.0	67.2	34.7	8.9	30.3	n
yolov8n_576x576	gray	37.1	51.9	20.1	6.4	30.3
-
yolov8s_320x320	color	53.6	70.7	33.2	5.6	11.8	l
yolov8s_320x320	gray	43.7	60.6	24.1	2.9	11.8	f
yolov8s_480x480	color	61.3	75.4	45.3	13.2	34.9	m
yolov8s_480x480	gray	52.4	67.9	34.9	7.5	34.9
-
yolov8m_320x320	color	60.6	76.5	41.8	8.8	20.7	l
yolov8m_320x320	gray	52.8	69.8	32.3	5.1	20.7	f
-
yolov9t_320x320	color	48.7	67.5	25.4	2.7	9.3	o
yolov9t_320x320	gray	37.6	55.0	17.8	1.6	9.3	h
yolov9t_480x480	color	56.0	72.1	37.7	9.2	23.5	p
yolov9t_480x480	gray	42.4	59.0	24.2	4.6	23.5	i
yolov9t_576x576	color	57.8	72.4	41.1	12.1	38.2	q
yolov9t_576x576	gray	43.6	59.3	28.0	7.0	38.2
-
yolov9s_320x320	color	55.5	73.1	34.6	6.2	12.4	o
yolov9s_320x320	gray	45.4	63.6	23.9	3.7	12.4	h
yolov9s_480x480	color	62.4	76.4	46.3	18.1	43.0	p
yolov9s_480x480	gray	53.7	69.9	34.6	9.2	43.0
-
yolov9m_320x320	color	60.2	76.4	41.4	10.2	39.1	o
yolov9m_320x320	gray	49.8	66.3	30.9	5.0	39.1
-
yolo11n_320x320	color	43.2	60.3	22.8	3.8	22.2	r
yolo11n_320x320	gray	31.7	46.8	14.3	0.6	22.2
-
cctv1(mobiledet)	color	46.6	64.9	24.6	1.8	1.5	s
cctv1(mobiledet)	gray	44.1	62.9	21.7	1.2	1.5	j
-
cctv3.3(mobiledet)	color	45.6	63.5	22.5	1.8	1.4	t
cctv3.3(mobiledet)	gray	45.5	63.6	22.3	2.0	1.4	k
-
cctv4(mobiledet)	color	45.5	63.3	23.5	2.2	1.5	u
cctv4(mobiledet)	gray	45.7	63.3	23.6	2.4	1.5	l

RPI5 4GB with USB accelerator

TFLITE

MODEL	mAP	L	M	S	RPI5
mobiledet	color	45.2	62.4	23.9	4.7
mobiledet	gray	40.2	57.7	17.5	3.0
-
cctv1	color	46.7	64.9	24.6	1.9
cctv1	gray	44.1	62.8	21.7	1.3
-
cctv3.3	color	45.5	63.4	22.3	1.8
cctv3.3	gray	45.6	63.6	22.4	2.0
-
cctv4	color	45.5	63.3	23.5	2.2
cctv4	gray	45.6	63.2	23.5	2.4

Models

The models are based on MobileDet with additional training data added.

CCTV1

Proof of concept model. Can only detect person. Resolution changed from 320x320 to 420x280. Achieves an improvement of 1.7mAP on gray images. Though at the cost of approximately 18% slower inference time.

CCTV2 failed run

Frigate compatible model. Can only detect `person`. Resolution changed from `320x320` to `340x340`. Approximately 18% slower inference time. Failed run, accuracy is worse than original model.

CCTV3

Frigate compatible model. Only trained on gray images. Can only detect person. Resolution changed from 320x320 to 340x340. Approximately 18% slower inference speed compared to mobiledet.

CCTV4

Frigate compatible model. Only trained on gray images. Can only detect person. Resolution changed from 320x320 to 340x340. Higher accuracy compared to mobiledet, but much slower inference speed. First model trained without transfer learning.

mAP is the average accuracy. L, M, S is the accuracy for Large, medium, and small boxes.

BASE MODELS

SIZE	NAME	FULL NAME
320x320	mobiledet	ssdlite_mobiledet_edgetpu/fp32

EVAL SETS

COCO Minival

Blacklisted images are removed from all testsets.

IMAGES	NAME	DESCRIPTION
720	coco-minival_person	Minival color and gray, person only.
349	color_coco-minival_person	Minival original, cctv1 labels. Gray images removed.
371	gray_coco-minival_person	Minival coverted to black and white, person only.

objects365-val-cctv-v1

All images are fairly easy with minimum ambiguity. Count represent the number of images with one or more of the object class.

CLASS	COUNT
total	1269
person	684
car	186
elephant	118
giraffe	105
bird	101
dog	101
bicycle	99
cat	93
zebra	80
horse	77
motorcycle	77
train	70
boat	65
cow	21
airplane	19
sheep	16

TRAINSETS

SIZE	NAME	DESCRIPTION
144k	cctv1	person only. Combination of coco2017 and voc2012. Blacklist not applied
144k	cctv2	Same as cctv1.
59k	cctv3	person only. coco2017 and voc2012 images coverted to gray.
54k	cctv3.1	The first 100k person IDs sorted.
50k	cctv3.2	The first 200k person IDs sorted.
44k	cctv3.3	The first 300k person IDs sorted.
21k	cctv4	All COCO person images sorted.

Naming convention.

type	description
color	Original images.
gray	Images converted to black and white.
empty	Combined color and gray.

DATASETS