SSD: Single Shot MultiBox Detector Wei Liu, Dragomir Anguelov, - - PowerPoint PPT Presentation

SSD: Single Shot MultiBox Detector

Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg Slides by: Sulabh Shrestha

Receptive Field

Ref: https://cv-tricks.com/object-detection/single-shot-multibox-detector-ssd/

▪ Deep feature maps

▪ Smaller size ▪ Larger receptive fields ▪ May miss small objects

▪ Shallow feature maps

▪ Larger size ▪ Smaller receptive fields ▪ May not be able to see larger objects

▪ Use multiple for corresponding receptive field sized objects

Use multiple

Architecture

▪ Base Network + Extra Feature Layer ▪ No FC layer ▪ Specific feature maps responsive to particular scale of objects

▪ Not necessarily same as the receptive field ▪ A hyper-parameter ▪ Dependent on data

8x8 Feature map 4x4 Feature map

VGG

Base Network

▪VGG 16 ▪Pool5 changed:

▪ 3x3 kernel instead of 2x2 ▪ Stride 1 instead of 2

▪1st 2 FCs replaced by CNN

▪ DeepLab LargeFOV

▪Last FC removed altogether ▪No dropouts used ▪Conv4_3 also used for prediction

▪ 4th Group of Conv ▪ 3rd kernel

Ref: VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION

Multiple Default Boxes

▪Similar to Anchor boxes of Faster-RCNN ▪Example feature map:

▪m x n ▪p-channels

▪For each location (i, j)

▪ Multiple default boxes (k) ▪ 3 x 3 x p-channel CNN for each box

▪ Confidence of each class, ci ; i Є [1, C] ▪ x, y, w, h ▪ (C+4) outputs

▪ Total outputs for 1 feature map:

▪ m * n * k * (#classes + 4)

m n p

Scale and Aspect ratio

▪ How many default boxes per location? ▪ Scale

▪ Related to but not exact as the receptive field ▪ If m feature maps used for prediction: ▪ smin = 0.2 ▪ smax = 0.9 ▪ Eg.

▪ s = 0.2 ▪ img-size = 300 ▪ Default box corresponding size = 0.2 * 300 = 60

▪ Aspect ratios(ar)

▪ {1, 2, 3, 1/2, 1/3} ~ k ▪ Width (wk

a) = sk √ ar

▪ Height (hk

a) = sk / √ ar

▪ Eg.

▪ s = 0.2, img-size = 300 ▪ ar = 1

• ->

w = 0.2 * 300 = 60 h = 0.2 * 300 = 60 ▪ ar = 2

• ->

w = 0.2 * √ 2 * 300 = 85 h = 0.2 / √ 2 * 300 = 42 ▪ ar = 1/2 --> w = 0.2 * √ ½ * 300 = 42 h = 0.2 / √ ½ * 300 = 85

Training

• Basenet pre-trained on ImageNet CLS-LOC dataset
• Fine-tuned for respective dataset
• Matching Strategy
• Any 𝐽𝑃𝑉𝑒𝑓𝑔𝑏𝑣𝑚𝑢𝑐𝑝𝑦

𝑕𝑠𝑝𝑣𝑜𝑒𝑢𝑠𝑣𝑢ℎ > 0.5 → 𝑞𝑝𝑡𝑗𝑢𝑗𝑤𝑓

• Simplifies learning problem
• Can detect object in multiple overlapping default boxes
• Loss
• Confidence loss (c)
• Softmax loss over multiple classes
• Localization loss (xywh)
• Smooth L1 loss
• Ground truth box(g) vs Default box(l)

Ref: https://github.com/rbgirshick/py-faster-rcnn/files/764206/SmoothL1Loss.1.pdf

Results

PASCAL VOC2007 test detection results PASCAL VOC2012 test detection results

Inference

• Filter boxes with low confidence
• NMS with 0.45 IOU
• Take top 200 detections
• Better mAP
• Faster FPS

VOC2007 Test data

Analysis

• Better than 2 stage network:
• Single network for localization and classification
• Better than YOLO
• Use multiple feature maps
• Use many more default boxes
• No FC layer
• Faster inference
• Fewer parameters
• Smaller input size
• Faster RCNN
• 600 min. size
• YOLO
• 448 x 448

Ablation Studies - 1

• Data Augmentation helps
• Original image
• Random sample of patch
• Sample patch
• IOUmin is 0.1, 0.3, 0.5, 0.7, 0.9
• More Multiple boxes helps
• Using FC instead of CNN

(Atrous)

• Similar result
• 20% slow

Ablation Studies - 2

• Use different number of feature maps
• Similar # of default boxes to make it fair
• More feature maps better
• Up to a certain extent
• Not using boundary defaults boxes better
• Avoid default boxes lying outside the image

Thank you

Questions?