Object Detection (coco-ssd)

line

Open a new browser window (or tab) and Copy / Paste(Ctrl+V) images from Internet. You can copy images from your PC e.g. by Photos or Paint. lite_mobilenet_v2 is smallest in size, and fastest in inference speed. mobilenet_v2 has the highest classification accuracy. Set new line width of boundary boxes. Canvas size corresponds to the expected by COCO-SSD image size (300x300 pixels). See console for detailes. See also Animated Fruits Detection (coco-ssd).

Similar applications:

With "upload image" and "webcam" dopelemon.me by Dhruv Jawalkar.
You can edit codepen and glitch demos with "fixed" images and webcam by Jason Mayes.

ToDo..?

I'd like to detect Roses.
Simple detector: lite MobileNet - "head" layers + big dense top layer (isn't it YOLO 1 :)? Isn't it similar to the "face.api" approach?
Retrain coco-ssd?
Training a YOLOv3 Object Detection Model with a Custom Dataset Jan 9

Comments

See github.com/tensorflow/tfjs-models/tree/master/coco-ssd.
Note that SSD model rescales image to "the expected 300x300 pixels (TF-lite)".

  // Warmup the model.
  const result = await this.model.executeAsync(tf.zeros([1, 300, 300, 3])) as tf.Tensor[];

The COCO dataset puts 80 class names in 90 class index. See Object Detection: class number 12 is missing in MSCoco.

TFjs notes updated 19 Jan 2020