This code is submitted for CV project. This is a Keras port of the SSD model architecture introduced by Wei Liu et al. in the paper SSD: Single Shot MultiBox Detector.
The original code was cloned from https://github.com/pierluigiferrari/ssd_keras
- Python 3.x
- Numpy
- TensorFlow 1.x
- Keras 2.x
- OpenCV
- Beautiful Soup 4.x
The code can train and evaluate several models on several datasets, I have used the models to train and evaluate parts of ICUB Dataset, each file contains a self explanatory comments(both the cloned files and the files that I have added)
the following Google Drive folder conatins all the supplementary materials ( Models, Dataset parts, test results, graphs) https://drive.google.com/open?id=1zTIgn_hta_T8plfcmsXSUBYTUD7tvIlK
I provide all the models I have trained using both Nivida Quardo P5000 and Nvidia GTX 950M: SSD300_ICUB_7, SSD512_COCO_ICUB_6(fine tuned), SSD7_ICUB_6,SSD7_ICUB_4,SSD7_ICUB_20_INSTANCE_DETECTION as well as MS_COCO resampled weights used to initialize ssd512 training, and VGG-16 feature extraction weights trained on ImageNet.
the described (in the document) datasets, which contain the training images and annotations with the list of train/val, along with the test images and annotations with the list. most dataset parts contain two test modalities: using unseen images from seen instances, and images from unseen instances. The parts were generated by two notebooks: batch copy images [category|instance] detection
are npz files generated by python scripts, containing the test results. the notebook: test_result_viewer can be used to re-generate mAP tables and Precision-recall graphs for these results.
the graphs generated for the tests in a jpg format