B0-4Y : EfficientNetB0 + 4 YOLO Layers + BiFPN on darknet AlexeyAB

EfficientNetB0 for classifier(AKA Backbone) and 4 YOLO Layers + BiFPN for detector on darknet Framework.

• enet4y2bifpn with 4.364 to calculate convolution layers
• enet4y2bifpn has Bidirectional Feature Pylamid Network
• 20MB weights file size
• 20days @ V100 GPU for taining with COCO
• 49.5 mAP50 with COCO 5K
• 81.0 mAP50 with VOC2007 test(as additional information)

enet4y2bifpn network structure

The original cfg which on AlexeyAB includes 2 YOLO Layers as detector.
enet4y2bifpn and the original enet-coco have same backbone classifier therefore use same backbone weights.

original README

Object Detector setup from Dataset download until estimation mAP

Download Dataset

$ mkdir work;cd work
$ ./scripts/get_coco_dataset.sh

Training using backbone weights pre-trained via ImageNet

backbone weigts
This weights of classifier may be generated for Cross Stage Partial Networks.

enet4y2bifpn-coco weights
This weights of detector is generated for enet4y2bifpn-coco.cfg.

• write coco.data

classes = 80
train   = work/coco/trainvalno5k.txt
valid   = work/coco/5k.txt
names   = data/coco.names
backup  = backup4y2

• download backbone weights 'enetb0_final.weights' from Website

• start training
  using 24batch size.

$ ./darknet detector train coco.data cfg/enet4y2bifpn-coco.cfg enetb0_final.weights  

Unfortunately wait 20days even if V100 GPGPU.

enet4y2bifpn-coco estimation mAP against enet-coco(detector as reference)

Notice:using 64batch size

Summary via pycocotools

/	AP50-95	AP50	AP75	Small	Medium	Large	Judgement
enet-coco	0.240	0.452	0.231	0.086	0.260	0.369	criterion
enet4y2-coco	0.236	0.471	0.213	0.091	0.258	0.339	target
enet4y2bifpn-coco	0.255	0.495	0.241	0.099	0.279	0.363	target
	win	win	win	win	win	lose	result

enet4y2bifpn-coco.cfg pycocotools details

$ ./darknet detector map coco.data cfg/enet4y2bifpn-coco.cfg backup4y2/enet4y2bifpn-coco-bigBatch_final.weights
$ ./scripts/pycocoeval.py -r results/coco_results.json -g work2/coco/annotations/instances_val2014.json

0.25474642780790624
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.255
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.495
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.241
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.099
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.279
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.363
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.229
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.354
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.377
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.189
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.409
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.515

enet4y2bifpn-voc estimation mAP as additional information

enet4y2bifpn-voc weights

Summary via darknet map

/	AP50	Comments
enet-voc	0.804	criterion
enet4y2-voc	0.792	target
enet4y2bifpn-voc	0.810	target

./darknet detector map voc.data cfg/enet4y2bifpn-voc.cfg backup4y2VOC/enet4y2-voc_best.weights
class_id = 0, name = aeroplane, ap = 89.30%      (TP = 235, FP = 31)
class_id = 1, name = bicycle, ap = 89.56%        (TP = 277, FP = 48)
class_id = 2, name = bird, ap = 80.07%           (TP = 350, FP = 81)
class_id = 3, name = boat, ap = 70.80%           (TP = 181, FP = 94)
class_id = 4, name = bottle, ap = 68.20%         (TP = 314, FP = 159)
class_id = 5, name = bus, ap = 90.40%            (TP = 182, FP = 41)
class_id = 6, name = car, ap = 89.76%            (TP = 1039, FP = 232)
class_id = 7, name = cat, ap = 89.56%            (TP = 302, FP = 48)
class_id = 8, name = chair, ap = 64.18%          (TP = 511, FP = 441)
class_id = 9, name = cow, ap = 84.06%            (TP = 207, FP = 72)
class_id = 10, name = diningtable, ap = 71.67%           (TP = 156, FP = 102)
class_id = 11, name = dog, ap = 85.62%           (TP = 398, FP = 111)
class_id = 12, name = horse, ap = 90.50%         (TP = 303, FP = 59)
class_id = 13, name = motorbike, ap = 89.71%     (TP = 272, FP = 73)
class_id = 14, name = person, ap = 85.57%        (TP = 3635, FP = 804)
class_id = 15, name = pottedplant, ap = 52.73%           (TP = 252, FP = 178)
class_id = 16, name = sheep, ap = 84.67%         (TP = 203, FP = 49)
class_id = 17, name = sofa, ap = 75.92%          (TP = 191, FP = 125)
class_id = 18, name = train, ap = 86.33%         (TP = 234, FP = 56)
class_id = 19, name = tvmonitor, ap = 81.50%     (TP = 249, FP = 74)

 for conf_thresh = 0.25, precision = 0.77, recall = 0.79, F1-score = 0.78
 for conf_thresh = 0.25, TP = 9491, FP = 2878, FN = 2541, average IoU = 59.83 %

 IoU threshold = 50 %, used Area-Under-Curve for each unique Recall
 mean average precision (mAP@0.50) = 0.810054, or 81.01 %

FPN Version, read next

Getting start from scratch without pre-trained backbone, read next

Convertion from darknet to pytorch and tensorflow next