This paper presents a general one-shot object localization algorithm called OneLoc. Current one-shot object localization or detection methods either rely on a slow exhaustive feature matching process or lack the ability to generalize to novel objects. In contrast, our proposed OneLoc algorithm efficiently finds the object center and bounding box size by a special voting scheme. To keep our method scale-invariant, only unit center offset directions and relative sizes are estimated. A novel dense equalized voting module is proposed to better locate small texture-less objects. Experiments show that the proposed method achieves state-of-the-art overall performance on two datasets: OnePose dataset and LINEMOD dataset. In addition, our method can also achieve one-shot multi-instance detection and non-rigid object localization.
We provide a demo, detecting a novel toy bear. The reference video is stored in data/ref.mp4, and the test sequence is stored in data/test.mp4.
To run the code, install the necessary packages and modify L27 of helper_math.cuh to point to the correct location of CUDA.
Then run,
python kp_server.py
and run
python train.py # in another terminal
After training for about 10 epochs, you will see the following detection results:
Of course, you can capture your own object, and replace ref.mp4/test.mp4. The only requirement is to ensure that the object occupies the whole image in ref.mp4; if not, you may have to annotate the bounding box and modify L42 of dataset.py to generate the correct voting target. Have fun!
Coming soon.