pred_hoi_dets.md

Description of pred_hoi_dets.hdf5 file

This is the format in which the hoi detections from any model need to be saved in order to use exp/hico_eval/compute_map.sh/py files for mAP evaluation.

HDF5 Directory Structure

.
+-- global_id1
|   +-- human_obj_boxes_scores
|   +-- start_end_ids
+-- global_id2
|   +-- human_obj_boxes_scores
|   +-- start_end_ids
...

HDF5 datasets description

human_obj_boxes_scores is a Nx9 (=4+4+1) dimensional numpy array with each row containing the box coordinates (x1,y1,x2,y2) for the human and object boxes, and score for the predicted hoi class.

start_end_ids is a 600x2 dimensional numpy array with i^th row containing the start and end row numbers in box_scores_rpn_ids for i^th class in the list of hico classes. Since the hoi class ids begin with '001', for an hoi class with id '006', the set of detections for a given global_id are obtained by

f = h5py.File(pred_dets_hdf5_path,'r')
hoi_id = '006'
start_id, end_id = f[global_id]['start_end_ids'][int(hoi_id)-1]
hoi_dets = f[global_id]['human_obj_boxes_scores'][start_id:end_id]
human_boxes = hoi_dets[:,:4]
object_boxes = hoi_dets[:,4:8]
scores = hoi_dets[:,8]