Anchor-free based object detection methods have enabled looking towards a new paradigm of object detection where a bounding box should not be constrained to 4-corner points. In the box ground-truth instances, it is estimated that about 40% of the area inside a box is comprised of background information; with rotated objects and non-rectangular objects in the higer end of this statistic. In this Master's Thesis, I try to mitigate that problem by deriving the convex hull of all the objects from their corresponding segmentation masks and constraining them to 8 points. These 8 point polygons with their centroids serve as the ground-truth for training the modified FCOS. Instead of the original FCOS where each valid location regresses for the left, right, top and bottom offsets of the ground-truth, our network regresses for the x and y offsets of each of the 8 arbitrary points. Figure below shows a prediction example from the MSCOCO dataset.
 |
|---|
| Convex FCOS Prediction |
All annotations were created from the MSCOCO_'17_ dataset. The folder Custom_Annotations contains the files to convert the COCO segmentation masks to 8-point Convex Hulls.
 |
|---|
| The original convex hull (in Yellow) and the 8-point Hull |
Currently, only ResNet-50 as a backbone has been tested. To train, make sure you have set the correct .npy file containing the convex hull annotations in Convex_Object_detection/fcos_core/data/datasets/coco.py and the correct path to the original COCO json annotations in Convex_Object_detection/fcos_core/config/paths_catalog.py. To train, use:
python3 tools/train_net.py --config-file configs/fcos/fcos_R_50_FPN_1x.yaml --skip-test DATALOADER.NUM_WORKERS 6 OUTPUT_DIR <save path>
python3 tools/test_net.py --config-file configs/fcos/fcos_R_50_FPN_1x.yaml MODEL.WEIGHT /MSCOCO/no_convexity_5C/<MODEL NAME>.pth TEST.IMS_PER_BATCH <BATCH SIZE>