fvilmos / centernet_tiny

Resource optimal implementation of Centernet in Tensorflow-keras, for multi object detection.

Centernet tiny

Centernet [1] is a point-based detector, that works without anchor-box proposals. This type of implementation eliminates the post-processing part known from Single Shot MultiBox Detectors [2], which usually consumes a lot of resources, therefore the Non-Maximum Suppression can be simplified.

With this approach, the object's center points and other properties (size, bounding box offset, depth, orientation, or others) are regressed, providing a large versatility of detection tasks.

This implementation provides an optimized version with low computational complexity, that can be easily deployed on embedded devices.

NOTE: This is an unofficial implementation of Centernet, for the original one consult the article [1]

Architecture

The architecture has the following parts:

1. Backbone - that can be adjusted to the computational complexity needed. currently, slim U-NET [3] is used to generate the necessary features maps
2. Centernet head - that generates Nr of classes + 2 (center point offset) + 2 (width and height), head width, head height heatmaps. Easily can be extended to regress for other type of values, i.e. 3D object detection or human pose detection

The implementation has a sample generator, that generates a noisy background and an object with different colors in the foreground, like circles, and rectangles.

Sample generator

Results

Measurements:
  Total params: 1,029,985
  Backbone computational complexity [FLOPS]: 0.263 G
  Backbone + head computational complexity [FLOPS]: 0.581 G
  IOU: 92.55 %
  Class Accuracy: 100.00 %, 

1. Detection results, 2. Centerpoint, 3. Width resgression values, 4. Height regression values

Resources

1. Objects as Points - arXiv:1904.07850 [cs.CV]
2. SSD: Single Shot MultiBox Detector - arXiv:1512.02325 [cs.CV]
3. U-Net: Convolutional Networks for Biomedical Image Segmentation - arXiv:1505.04597 [cs.CV]
4. Focal Loss for Dense Object Detection

/Enjoy.