YOLOv2

28 July 2020

Photo by Grace Brauteseth

Origin: YOLO9000: Better, Faster, Stronger

Imporvements comparing to YOLOv1

1. Add BN

2. Hight Resolution classifier

a.  Train on ImageNet (224 x 224)
b.  Resize & Finetune on ImageNet (448 x 448)
c.  Finetune on dataset
d.  Get 13 x 13 grid finally (7 x 7 grid before)

3. Use Anchors

4. Fine-Grained Features

a.  Lower features are concatenated directly to heigher features
b.  A new layer is added for that purpose: reorg

[abcdefghijklmnop]===>[acik][bdjl][fhnp][egmn]

5. Multi-Scale Training

• Remove FC layers: Can accpet any size of inputs, enhance model robustness.
• Size across 320, 352, …, 608. Change 10 per epochs
       [border % 32 = 0, decided by down sampling

Anchor in YOLOv2

Box Generation	#	Avg IOU
Cluster SSE	5	58.7
Cluster IOU	5	58.7
Anchor Boxes	9	58.7
Cluster IOU	9	58.7

1. Anchor size and number

a.  Faster RCNN: 9 by hands
b.  YOLOv2: 5 by K-Means [dist: 1 − IOU(bbox, cluster)]

2. Anchors, Truth BBoxes & Predicted bboxes

Anchors: 0.57273, 0.677385, …, 9.77052, 9.16828

               [10 numbers: (awi,ahi) * 5)]
               anchors[0] = awi = awiW * 13

Truth Anchor:

1. oringinal bbox: [xo,yo,wo,ho]∈[0,W∣H]
2. normalize original bbox: [xr,yr,wr,hr]∈[0,1]
   • [xr,yr,wr,hr]=[xo/W,yo/H,wo/W,ho/H]
3. Transfer to 13 x 13 grid and box: [xs,ys,ws,hs]∈[0,13]
   • [xi,yi,wi,hi]=[xr,yr,wr,hr]∗(13∣13)
   • save this for calculating
   • transfer to 0~1 corresponding to each grid cell
4. final box: [xf,yf,wf,hf]∈[0,1]
   • xf=xi−i     // i,j = 13 x 13 grid
   • yf=yi−j
   • Wf=log(W/anchors[0])
   • Hf=log(H/anchors[1])

Predicted Anchor:

bx=σ(tx)+cxby=σ(ty)+cybw=pwetwbh=phethPr( object )∗IOU(b, object )=σ(to)

The Model Darknet-19

Output of YOLOv2: [0: 25]

one grid cell: S2 * B * [x, y, w, h, C0,...,CN]

• detection layer
  • 3 x 331024 Conv
  • add a passthrough layer
  • 1*1 avg pooling

Loss Function

∑Wi=0∑Hj=0∑Ak=01Max IOU<Threshλnoobj∗(−boijk)2+1t<12800λprior∗∑r∈(x,y,w,h)(priorrk−brijk)2+1truthk(λcoord∗∑r∈(x,y,w,h)(truthr−brijk)+λobj∗(IoUktruth−boijk)2+λclass∗(∑Cc=1(truthc−bcijk)2))

• No longer use the square root
• Confidence: 1 convert to IoU

---

The path from YOLO to YOLOv2

	YOLO								YOLOv2
batch norm?		√	√	√	√	√	√	√	√
hi-res classifier?			√	√	√	√	√	√	√
convolutional?				√	√	√	√	√	√
anchor boxes?				√	√
new network?					√	√	√	√	√
dimension priors?						√	√	√	√
location prediction?						√	√	√	√
passthrough?							√	√	√
multi-scale?								√	√
hi-res detector?									√
VOC2007 mAP	63.4	65.8	69.5	69.2	69.6	74.4	75.4	76.8	78.6