eo_seals_yolo_tiny_3l_512_1to1bg_nounk

eo_seals_yolo_tiny_3l_512_1to1bg_nounk

Model Info
Input Size	512x512
Classes	Ringed Seal, Bearded Seal
Detection layers	P5/32, P4/16, P3/8
Iterations	~60,000
Batch Size	64
Repo	AlexyAB/darknet b5ff7f4
Initialization	yolov3-tiny.conv.15
Training Net	eo_seals_yolo_tiny_3l_512_1to1bg_nounk_TRAIN.cfg
Inference Net	eo_seals_yolo_tiny_3l_512_1to1bg_nounk.cfg

Notes: The reason we have a train net and an inference net is so that model is compatible with pjreddie's darknet. Class distribution scaling and other training-time improvements were used that are not compatible with the older darknet network parser.

Dataset Info
Background Ratio Train	~1:1
Background Ratio Test	~1:1

Benchmark speed (1080ti,batch_size=1):

Results @ .25 IOU, Thresh = .25	Ringed Seal	Bearded Seal	Combined
TP	2645	283	2928
FP	423	141	564
FN	200	10	210
Precision	0.87	.67	0.84
Recall	0.93	.93	.93

Model Performance

Train Set	Test Set