0

I'm working on object detection of various sorts of animals using the Tensorflow Object Detection API. In the past I successfully applied MobileNet v1 to various settings and I used to be happy with the results.

Now, I encountered a problem with a new species that is about 1/3 smaller than animals I dealt with before. Visually, the animals look the same up to a scale, meaning that the bounding boxes to be predicted are rather in the range of 5-15% of the image size than 20%-30% as before.

I have the feeling there should be some hyperparameter I need to tweak in order to get stuff back to working, but I struggle to find the right one the pipeline config. I already experimented with tuning min_scale and max_scale of the anchor_generator towards smaller values, but with no success.

Interestingly, using Faster RCNN works right away on the exact same data.

Any ideas what could be tried?

zimmermc
  • 623
  • 2
  • 6
  • 13
  • Forgot to mention the exact models I work with: What doesn't work is ssd_mobilenet_v2_coco_2018_03_29 compared to a faster_rcnn_resnet50_coco_2018_01_28 which performs well. – zimmermc Jun 25 '19 at 07:40
  • Faster R-CNN and SSD are both object detectors while mobilenet is a feature extractor. The problem is SSD's weak ability to detect small objects. But you can try this [solution](https://stackoverflow.com/questions/56071683/ssd-mobilenet-model-does-not-detect-objects-at-longer-distances.) to see if it works. – Danny Fang Jun 25 '19 at 09:59

1 Answers1

0

For smaller objects, you can try RetinaNet/SSD-resnet-FPN. The FPN (Feature Pyramid Networks) should work better than mobilenet.

stormzhou
  • 316
  • 2
  • 5