def get_org_gt_dets():
dets_annots_path = get_dets_annots_path(data_dir)
dets_arr = load_dets_annots(dets_annots_path)
frm_ids = dets_arr[:, 0].astype(int)
obj_ids = dets_arr[:, -1].astype(int)
coords = dets_arr[:, 1:5]
boxes = coords_to_boxes(coords)
return frm_ids, obj_ids, boxes
def amend_dets(masks, in_dets):
out_dets = in_dets
in_coords = in_dets[:, 2:6]
in_boxes = coords_to_boxes(in_coords)
out_boxes = boxes_align_masks(masks, in_boxes)
out_coords = boxes_to_coords(out_boxes)
out_dets[:, 2:6] = out_coords
return out_dets
def parse_gt_annots():
gt_masks_tensor = load_masks_new()
gt_dets = load_dets_new()
## parse dets
gt_frm_ids = gt_dets[:, 0].astype(int)
gt_obj_ids = gt_dets[:, 1].astype(int)
gt_coords = gt_dets[:, 2:6]
gt_boxes = coords_to_boxes(gt_coords) ## tight_gt_boxes
return gt_frm_ids, gt_obj_ids, gt_boxes, gt_masks_tensor
def get_gt_dets():
dets_annots_path = get_dets_annots_path(data_dir)
dets_arr = load_dets_annots(dets_annots_path)
if debug_flag:
print 'dets_annots_path:', dets_annots_path
print 'dets_arr.shape:', dets_arr.shape
frm_ids = dets_arr[:, 0].astype(int)
obj_ids = dets_arr[:, -1].astype(int)
coords = dets_arr[:, 1:5]
boxes = coords_to_boxes(coords)
return frm_ids, obj_ids, boxes
def interploate_dets(disc_dets_path, cont_dets_path):
print '==================interploate det proposals=========================='
print '--------------------------discrete_det_arr--------------------------------------------'
disc_det_arr = load_txt_to_fltarr(disc_dets_path)
#print 'disc_det_arr:\n', disc_det_arr
tmp_det_arr = disc_det_arr # frm-no, obj_id, x1,y1,w,h
tmp_coords = tmp_det_arr[:, 2:6] ## x1,y1,w,h
tmp_boxes = coords_to_boxes(tmp_coords) ## x1,y1,x2, y2
tmp_det_arr[:, 2:6] = tmp_boxes
print '--------------------------------------------------------------------'
#frm_no,obj_id,x1,y1,x2,y2
col = 1 ## obj_id (col)
sort_det_arr = tmp_det_arr[np.argsort(
tmp_det_arr[:, col])] ## sort according to obj_id
obj_ids = np.unique(sort_det_arr[:, 1]).astype(int)
print '----------------------------cont_det_arr-------------------------------------------'
cont_det_arr = []
tmp_cont_det_arr = np.zeros((0, sort_det_arr.shape[1]))
## interploate box based on each obj_id
for obj_id in obj_ids:
#print '--------------------------obj_id=', obj_id, '------------------------------------------------'
row_indexes = np.where(sort_det_arr[:, 1] == obj_id)
disc_label_det_arr = sort_det_arr[row_indexes]
cont_label_det_arr = interploate_label_dets(obj_id, disc_label_det_arr)
# print 'disc_label_det_arr.shape:', disc_label_det_arr.shape
# print 'cont_label_det_arr.shape:', cont_label_det_arr.shape
if disc_label_det_arr.shape == cont_label_det_arr.shape:
print 'obj_id that equal:', obj_id
cont_label_boxes = cont_label_det_arr[:, 2:
6] ## convert back from boxes to coords
cont_label_coords = boxes_to_coords(cont_label_boxes)
cont_label_det_arr[:, 2:6] = cont_label_coords
tmp_cont_det_arr = np.vstack((tmp_cont_det_arr, cont_label_det_arr))
cont_det_arr = np.asarray(tmp_cont_det_arr) #convert list to array
write_detarr_as_mot(cont_dets_path,
cont_det_arr) ## save in local disk as .txt
if debug_flag:
print 'cont_det_arr.shape:', cont_det_arr.shape
def test_vis_gt_boxes():
imgset_path, jpg_dir, im_ext = io.get_ims_info()
im_names = ld.get_im_names(imgset_path)
ims = ld.load_color_images(jpg_dir, im_names, im_ext)
data_dir = os.path.join(mcfg.DATA.DATA_DIR, set_name)
tud_dets_gt_path = get_dets_annots_path(data_dir)
if debug_flag:
print 'imgset_path:', imgset_path
print 'jpg_dir:', jpg_dir
print 'im_ext:', im_ext
print 'data_dir:', data_dir
print 'tud_dets_gt_path:', tud_dets_gt_path
tud_dets_arr = load_dets_annots(tud_dets_gt_path)
frm_ids = tud_dets_arr[:, 0].astype(int)
obj_ids = tud_dets_arr[:, -1].astype(int)
# uniq_obj_ids=np.sort(np.unique(obj_ids))
# for u_id in uniq_obj_ids:
# label_row_indexes=np.where(obj_ids==u_id)[0]
# label_frm_ids=frm_ids[label_row_indexes]
# print 'u_id:',u_id
# print 'label_frm_ids:', label_frm_ids
# if debug_flag:
# print 'tud_dets_arr.shape:', tud_dets_arr.shape
# print 'frm_ids:', frm_ids
# print 'obj_ids:', obj_ids
# print 'uniq_obj_ids:', uniq_obj_ids
coords = tud_dets_arr[:, 1:5]
##vis_arr=tud_dets_arr[:,5]/100.0 ##make vis range in [0, 1.0]
boxes = coords_to_boxes(coords) ## unnormal
det_dir = mcfg.PROPOSAL.RES_DIR.replace('Continious', 'Discrete')
gt_det_dir = os.path.join(det_dir, 'vis_gt_boxes')
for n_id in xrange(len(im_names)):
im = ims[n_id]
im_name = im_names[n_id]
frm_id = int(im_name.split('_')[1])
row_indexes = np.where(frm_ids == frm_id)[0]
frm_boxes = boxes[row_indexes]
frm_obj_ids = obj_ids[row_indexes]
vis_gt_bboxes(gt_det_dir, im, im_name, frm_boxes, frm_obj_ids)
def get_init_dets(cont_flag=False):
det_boxes = []
det_masks = []
det_obj_ids = []
det_frm_ids = []
bin_mask_dir = mcfg.MASK.BIN_LAB_DIR
init_mask_dir = os.path.join(bin_mask_dir, set_name)
prop_file_path, prop_file_ext = io.get_dets_info(set_name)
if not cont_flag:
init_mask_dir = init_mask_dir.replace('Continious', 'Discrete')
prop_file_path = prop_file_path.replace('Continious', 'Discrete')
prop_arr = ld.load_det_proposals(prop_file_path)
frm_ids = np.sort(np.unique(prop_arr[:, 0]).astype(int))
frm_num = len(frm_ids)
img_box = [0, 0, im_width - 1, im_height - 1] ## not used
## calculate based on per-frame
##for n_id in xrange(1):
for n_id in xrange(frm_num):
frm_id = frm_ids[n_id]
row_indexs = np.where(prop_arr[:, 0] == frm_id)[0] ##frame indexes
frm_prop_arr = prop_arr[row_indexs]
frm_labels = frm_prop_arr[:, 1].astype(int)
im_name = set_name + '_' + str(frm_ids[n_id] + shift_num).zfill(6)
frm_coords = frm_prop_arr[:, 2:6]
flt_frm_boxes = coords_to_boxes(frm_coords)
frm_masks = load_frame_masks(init_mask_dir, frm_labels, '', im_name)
int_frm_boxes = boxes_float_to_int(flt_frm_boxes, frm_masks)
##max_masks=get_max_masks(masks,boxes,img_box) ## not used
det_frm_ids.append(frm_id)
det_obj_ids.append(frm_labels)
det_boxes.append(int_frm_boxes)
det_masks.append(frm_masks)
# if debug_flag:
# print 'cont_flag:', cont_flag
# print 'init_mask_dir:', init_mask_dir
# print 'prop_file_path:', prop_file_path
# print 'prop_arr.shape:', prop_arr.shape
# print 'frm_num:', frm_num
return det_frm_ids, det_obj_ids, det_boxes, det_masks
def dets_tud_to_mot(tud_dets_path, mot_dets_path):
tud_dets_arr = load_dets_annots(tud_dets_path)
frm_ids = tud_dets_arr[:, 0].astype(int)
obj_ids = tud_dets_arr[:, -1].astype(int)
coords = tud_dets_arr[:, 1:5]
##vis_arr=tud_dets_arr[:,5]/100.0 ##make vis range in [0, 1.0]
boxes = coords_to_boxes(coords) ## unnormal
row_num = len(frm_ids)
mot_dets_arr = np.ones((row_num, 10)) * -1
mot_dets_arr[:, 0] = frm_ids
mot_dets_arr[:, 1] = obj_ids
mot_dets_arr[:, 2:6] = coords
mot_dets_arr[:, 6] = 1
##write_detarr_as_mot(mot_dets_path,mot_dets_arr) ## org
if debug_flag:
# print frm_ids.shape
# print obj_ids.shape
# print obj_ids
print mot_dets_arr
def get_enlarge_masks(track_mask_dir, cont_dets):
## mini_masks for ref-masks
mini_masks = get_ref_masks(track_mask_dir, cont_dets)
# transfer small masks(same size with boxes) to large masks(same size with color image)
mask_num = len(mini_masks)
masks = np.zeros((mask_num, im_height, im_width), dtype=int)
## mini_boxes
dets = get_ref_dets(track_mask_dir, cont_dets)
coords = dets[:, 2:6]
boxes = coords_to_boxes(coords)
boxes = boxes.astype(int)
## max_boxes
max_boxes = np.zeros((mask_num, 4))
max_boxes[:, 2] = im_width - 1
max_boxes[:, 3] = im_height - 1
max_boxes = max_boxes.astype(int)
## max_masks
max_masks = masks_align_boxes(mini_masks, boxes, max_boxes)
arr_masks = np.asarray(max_masks)
return arr_masks
def cal_per_label(rgb_ims,im_names,forward_flows,backward_flows,filt_obj_ids,gt_dets,gt_masks,corrup_boxes,corrup_masks):
print '===============calculate per label================='
## parameters
start_frm_id=1
for f_obj_id in filt_obj_ids:
print '---------------------------------------------',f_obj_id
gt_masks_dir,corrup_masks_dir,pick_masks_dir=create_masks_dirs(f_obj_id)
o_gt_dets,row_indexes=dets_filter_ids(gt_dets,[f_obj_id])
o_gt_frm_ids,o_gt_obj_ids,o_gt_boxes=parse_dets1(o_gt_dets)
o_gt_masks=gt_masks[row_indexes]
o_corrup_boxes=corrup_boxes[row_indexes]
o_corrup_masks=corrup_masks[row_indexes]
## filter color images
filt_indexes=o_gt_frm_ids-start_frm_id
im_arr=np.asarray(rgb_ims)
cur_ims=im_arr[filt_indexes]
cur_im_names=im_names[filt_indexes]
fw_flows=forward_flows[filt_indexes[:-1]]
bw_flows=backward_flows[filt_indexes[:-1]]
save_vis_masks(cur_ims,cur_im_names,gt_masks_dir,o_gt_masks,corrup_masks_dir,o_corrup_masks) #v1
if debug_flag:
print 'o_gt_dets:', o_gt_dets
print 'gt_masks_dir:',gt_masks_dir
print 'corrup_masks_dir:', corrup_masks_dir
print 'pick_masks_dir:', pick_masks_dir
print 'o_gt_frm_ids:', o_gt_frm_ids
#----------------------------------------------------------------------------------------
## prepare dataset for 'refine-masks' algorithm
# preprocessing corrup_masks and corrup_boxes(corrup_dets): make masks has the same size with boxes
o_corrup_boxes=o_corrup_boxes.astype(int)
tailor_masks=masks_align_boxes(o_corrup_masks,o_corrup_boxes)
tailor_masks=np.asarray(tailor_masks)
## prepare (corrup) dets
corrup_coords=boxes_to_coords(o_corrup_boxes)
corrup_dets=o_gt_dets
corrup_dets[:,2:6]=corrup_coords
# 6> refine masks
##input: prop_arr,masks,obj_ids, forward_flows, backward_flows
refine_masks(pick_masks_dir,rgb_ims,forward_flows,backward_flows,tailor_masks,corrup_dets)
ref_dets=get_ref_dets(pick_masks_dir,corrup_dets) ## verified
# ref_masks=get_ref_masks(pick_masks_dir,corrup_dets)
ref_masks=get_enlarge_masks(pick_masks_dir,corrup_dets)
ref_coords=ref_dets[:,2:6]
ref_boxes=coords_to_boxes(ref_coords)
verify_ref_annots(gt_frm_ids,gt_obj_ids,gt_boxes,gt_masks,ref_boxes,ref_masks,write_res_flag)