Models

General Structure

This directory holds the models, both those downloaded automatically from GluonCV the ones that are trained. Trained models will be put in a subdirectory named with the args.save-prefix argument. For example our structure looks as follows:

VidDet/models/README.md
VidDet/models/darknet53-2189ea49.params        <- downloaded by gluoncv
VidDet/models/mobilenet1.0-efbb2ca3.params     <- downloaded by gluoncv

VidDet/models/0001/                            <- our own trained model

GluonCV ModelZoo

GluonCV provides a few pre-trained models in their Model Zoo. Such models are downloaded automatically when specified in GluonCV with the appropriate gluoncv.model_zoo.get_model() function call, however we present these models for download and organise them similarly to our trained models.

GCV1 (0001 Alternative)

Download

Trained on PascalVOC trainval 07+12

yolo3_darknet53_voc

GCV2 (0003 Alternative)

Download

Trained on MSCoco train 17

yolo3_darknet53_coco

GCV3 (0006 Alternative)

Download

Uses MobileNet1.0 instead of DarkNet53

Trained on PascalVOC trainval 07+12

yolo3_mobilenet1.0_voc

GCV4 (0008 Alternative)

Download

Uses MobileNet1.0 instead of DarkNet53

Trained on MSCoco train 17

yolo3_mobilenet1.0_coco

Our Models

Our models, log files, and evaluation results are available for download by clicking on each model ID below.

0001

Download

Trained on PascalVOC trainval 07+12

python train_yolov3.py --dataset voc --gpus 0,1,2,3 --save_prefix 0001 --num_workers 16 --warmup_lr 0.0001 --warmup_epochs 3 --syncbn True

0002

Download (SOON)

Trained on ImageNetDET train_nonempty

python train_yolov3.py --dataset det --gpus 0,1,2,3 --save_prefix 0002 --num_workers 16 --warmup_lr 0.0001 --warmup_epochs 3 --syncbn True

0003

Download

Trained on MSCoco train 17

python train_yolov3.py --dataset coco --gpus 0,1,2,3 --save_prefix 0003 --num_workers 16 --warmup_lr 0.0001 --warmup_epochs 3 --syncbn True

0004

Download

Trained on ImageNetVID train17_ne_0.04

python train_yolov3.py --dataset vid --gpus 0,1,2,3 --save_prefix 0004 --num_workers 16 --warmup_lr 0.0001 --warmup_epochs 3 --syncbn True --frames 0.04

0006

Download

Uses MobileNet1.0 instead of DarkNet53

Trained on PascalVOC trainval 07+12

python train_yolov3.py --network mobilenet1.0 --dataset voc --gpus 0,1,2,3 --save_prefix 0001 --num_workers 16 --warmup_lr 0.0001 --warmup_epochs 3 --syncbn True

0007

Download (SOON)

Uses MobileNet1.0 instead of DarkNet53

Trained on ImageNetDET train_nonempty

python train_yolov3.py --network mobilenet1.0 --dataset det --gpus 0,1,2,3 --save_prefix 0002 --num_workers 16 --warmup_lr 0.0001 --warmup_epochs 3 --syncbn True

0008

Download (SOON)

Uses MobileNet1.0 instead of DarkNet53

Trained on MSCoco train 17

python train_yolov3.py --network mobilenet1.0 --dataset coco --gpus 0,1,2,3 --save_prefix 0003 --num_workers 16 --warmup_lr 0.0001 --warmup_epochs 3 --syncbn True

0009

Download (SOON)

Uses MobileNet1.0 instead of DarkNet53

Trained on ImageNetVID train17_ne_0.04

python train_yolov3.py --network mobilenet1_0 --dataset vid --gpus 0,1,2,3 --save_prefix 0004 --num_workers 16 --warmup_lr 0.0001 --warmup_epochs 3 --syncbn True --frames 0.04

Results

Evaluated with voc and coco metrics. Box Area's - Small <32, Medium 32-96, Large >96

Model	Trained On	Tested On	VOC12	AP.5-.95	AP.5	AP.75	APS	APM	APL
0001	VOC trainval 07+12	VOC test 07	.835	.463	.733	.510	.118	.317	.559
GCV1	VOC trainval 07+12	VOC test 07	.836	.462	.735	.500	.113	.304	.564
0006	VOC trainval 07+12	VOC test 07	.751	.356	.656	.346	.095	.205	.438
GCV3	VOC trainval 07+12	VOC test 07	.779	.396	.677	.418	.104	.245	.486
0003	COCO train 17	COCO val 17	.525	.288	.515	.296	.136	.306	.427
GCV2	COCO train 17	COCO val 17	.579	.360	.571	.387	.173	.387	.522
GCV4	COCO train 17	COCO val 17	.500	.286	.488	.299	.132	.298	.423
0004	VID train17_ne_0.04	VID val_ne_0.04	.478	.274	.453	.298	.031	.130	.330

Evaluated with vid metric. Box Area's - Small <50, Medium 50-150, Large >150. Instance's Speed (motion IoU) - SLow >0.9, MOderate 0.7-0.9, FAst <0.7

Model	Trained On	Tested On	mAP	APS	APM	APL	APSL	APMO	APFA
0004	VID train17_ne_0.04	VID val_ne_0.04	.454	.136	.328	.555	.522	.442	.292