Improving custom object detection false positives

System information

• What is the top-level directory of the model you are using: research/object_detection/
• Have I written custom code: No, I followed the tutorial
• OS Platform and Distribution: Ubuntu 16.04
• TensorFlow installed from (source or binary): source
• TensorFlow version (use command below): 1.14.0
• **Bazel version **: N/A
• CUDA/cuDNN version: 10.0
• GPU model and memory: NVIDIA Tesla V100
• Exact command to reproduce: N/A

Feature Recommendation

I am using the tensorflow custom detection api to detect custom furniture. I have found that I get very good accuracy when I see one of the classes. However, when seeing images that are not of any category, there is a high likelihood of seeing false positives.

I googled around looking for a solution (issue 29078, SO post, #2544). From my research, I could not find a definitive answer.

I was able to solve this problem myself and think that a new feature could solve this problem for others. My solution was to randomly feed in unlabeled images from the coco dataset to show my model negative examples. I believe that this could be a data augmentation option.

How do others feel about this? Thanks!