FabVis-Models

Clone the repository,

git clone https://github.com/IntellisenseLab/FabVis-Models.git

Download pretrained weight files from here or trained weight files from here and extract them and move the each folder into the repo folder.

Yolov4 Model Comparison

Comparison of Resolution

Evaluating the effect of resolution on the performance of the model

Model	Reso lution	Batch	Sub divisions	Pre trained	Precision	Recall	F1 score	mAP @0.5	Avg IoU	Output
Tiny	640	64	16	Yes	0.33	0.07	0.12	0.1400	0.2219	terminal
Tiny	416	64	16	Yes	0.36	0.10	0.16	0.1821	0.2474	terminal
Tiny	320	64	16	Yes	0.45	0.14	0.22	0.1781	0.3002	terminal
Tiny	256	64	16	Yes	0.33	0.09	0.15	0.1839	0.2221	terminal

Comparison of use of Weights

Evaluating the effect of resolution on the performance of the model

Model	Reso lution	Batch	Sub divisions	Pre trained	Precision	Recall	F1 score	mAP @0.5	Avg IoU	Output
Tiny	640	64	16	Yes	0.33	0.07	0.12	0.1400	0.2219	terminal
Tiny	640	64	16	No	0.43	0.10	0.17	0.1194	0.2881	terminal
Tiny	416	64	16	Yes	0.36	0.10	0.16	0.1821	0.2474	terminal
Tiny	416	64	16	No	0.45	0.10	0.16	0.1918	0.2978	terminal
Tiny	320	64	16	Yes	0.45	0.14	0.22	0.1781	0.3002	terminal
Tiny	320	64	16	No	0.43	0.12	0.18	0.1684	0.2948	terminal
Tiny	256	64	16	Yes	0.33	0.09	0.15	0.1839	0.2221	terminal
Tiny	256	64	16	No	0.33	0.10	0.15	0.1512	0.2194	terminal

Comparison of Models

All the values range between 0 - 1

Model	Reso lution	Batch	Sub divisions	Pre trained	Precision	Recall	F1 score	mAP @0.5	Avg IoU	Output
yolov4-tiny	416	16	8	Yes	0.21	0.08	0.11	0.1080	0.1422	terminal
yolov4-tiny	640	16	8	Yes	0.35	0.09	0.14	0.1512	0.2455	terminal
yolov4	416	16	8	Yes	0.62	0.33	0.43	0.1846	0.4690	terminal
yolov4	640	16	8	Yes	0.73	0.44	0.55	0.2617	0.5446	terminal
yolov4x-mish	416	16	8	Yes	0.65	0.34	0.44	0.2312	0.4881	terminal
yolov4x-mish	640	16	8	Yes	0.72	0.41	0.53	0.2178	0.5433	terminal
yolov4-csp	416	16	8	Yes	0.77	0.36	0.49	0.2328	0.5720	terminal
yolov4-csp	640	16	8	Yes	0.69	0.44	0.53	0.2070	0.5164	terminal
yolov4-csp-swish	416	16	8	Yes	0.68	0.33	0.44	0.1546	0.5771	terminal
yolov4-csp-swish	640	16	8	Yes	0.68	0.45	0.54	0.1464	0.5052	terminal
yolov4-csp-x-swish	416	16	8	Yes	0.68	0.33	0.45	0.0955	0.5172	terminal
yolov4-csp-x-swish	640	16	8	Yes	0.71	0.43	0.54	0.1798	0.5329	terminal
yolov4-p5	896	16	8	Yes	0.68	0.41	0.51	0.3619	0.5559	terminal
yolov4-p6	1280	16	16	Yes	0.60	0.43	0.50	0.3642	0.4763	terminal

Comparison of Batch size and subdivision size

All the values range between 0 - 1 Used Model is yolov4-tiny, and resolution is 416x416

Minibatch	Batch	Sub divisions	Pre trained	Precision	Recall	F1 score	mAP @0.5	Avg IoU	Output
2	16	8	Yes	0.21	0.08	0.11	0.1080	0.1422	terminal
4	64	16	Yes	0.36	0.10	0.16	0.1821	0.2474	terminal
8	64	8	Yes	0.34	0.11	0.16	0.1789	0.2176	terminal
4	128	32	Yes	0.30	0.10	0.15	0.1372	0.2079	terminal
8	128	16	Yes	0.36	0.11	0.17	0.1901	0.2387	terminal
16	128	8	Yes	0.29	0.09	0.14	0.1900	0.2015	terminal
4	256	64	Yes	0.35	0.12	0.18	0.1256	0.2395	terminal
8	256	32	Yes	0.30	0.11	0.16	0.1725	0.2068	terminal
16	256	16	Yes	0.31	0.09	0.14	0.1210	0.2170	terminal
32	256	8	Yes	0.32	0.10	0.16	0.1864	0.2205	terminal

Comparison of Final Model (Yolov4-p5 896x896)

Batch/Subdivision diference

Detection time confidence threshold variation

Detection time iou threshold variation

Detection time confidence and iou threshold variation

Comparison of resource utilization of the largest models

Idle utilization

yolov4-p5 utilization

• RAM Memory usage - ~1.8 GB
• GPU Memory usage - ~3.5 GB
• CPU count (used) - 1

yolov4-p6 utilization

• RAM Memory usage - ~2.2 GB
• GPU Memory usage - ~6.1 GB
• CPU count (used) - 1

Model Train Commands

Following commands will enable the training of models in coorindation of Preprocess system

Yolov4-Tiny

./darknet detector train ../config/obj.data ../FabVis-Models/config/yolov4-tiny.cfg ../FabVis-Models/preTrainedWeights/yolov4-tiny.conv.29 -dont_show -mjpeg_port 8090 -map

Yolov4

./darknet detector train ../config/obj.data ../FabVis-Models/config/yolov4.cfg ../FabVis-Models/preTrainedWeights/yolov4.conv.137 -dont_show -mjpeg_port 8090 -map

Yolov4x-mish

./darknet detector train ../config/obj.data ../FabVis-Models/config/yolov4x-mish.cfg ../FabVis-Models/preTrainedWeights/yolov4x-mish.conv.166 -dont_show -mjpeg_port 8090 -map

Yolov4-csp-swish

./darknet detector train ../config/obj.data ../FabVis-Models/config/yolov4-csp-swish.cfg ../FabVis-Models/preTrainedWeights/yolov4-csp-swish.conv.164 -dont_show -mjpeg_port 8090 -map

Yolov4-csp-x-swish

./darknet detector train ../config/obj.data ../FabVis-Models/config/yolov4-csp-x-swish.cfg ../FabVis-Models/preTrainedWeights/yolov4-csp-x-swish.conv.192 -dont_show -mjpeg_port 8090 -map

Yolov4-csp

./darknet detector train ../config/obj.data ../FabVis-Models/config/yolov4-csp.cfg ../FabVis-Models/preTrainedWeights/yolov4-csp.conv.142 -dont_show -mjpeg_port 8090 -map

Yolov4-p5

./darknet detector train ../config/obj.data ../FabVis-Models/config/yolov4-p5.cfg ../FabVis-Models/preTrainedWeights/yolov4-p5.conv.232 -dont_show -mjpeg_port 8090 -map

Yolov4-p6

./darknet detector train ../config/obj.data ../FabVis-Models/config/yolov4-p6.cfg ../FabVis-Models/preTrainedWeights/yolov4-p6.conv.289 -dont_show -mjpeg_port 8090 -map