def reshape(self, bottom, top):
#sample a category
categid = np.random.choice(self.numclasses)
#sample an image for this category
imid = self.data_percateg[categid]['imids'][np.random.choice(
len(self.data_percateg[categid]['imids']))]
img = cv2.imread(self._params.imgpath.format(self.names[imid]))
#get all possibilities for this category
start = self.data_percateg[categid]['im_end_index'][imid] + 1
stop = self.data_percateg[categid]['im_end_index'][imid + 1]
#pick a box
idx = np.random.choice(np.arange(start, stop + 1),
cfg.TRAIN_SAMPLES_PER_IMG)
boxid = self.data_percateg[categid]['boxids'][idx]
boxes = self.boxes[imid][boxid, :] - 1
instid = self.data_percateg[categid]['instids'][idx]
#load the gt
[inst, categories] = sbd.load_gt(self.names[imid])
masks = np.zeros((idx.size, 1, inst.shape[0], inst.shape[1]))
for k in range(idx.size):
masks[k, 0, :, :] = (inst == instid[k] + 1).astype(np.float32)
categids = categid * np.ones(idx.size)
#get the blobs
im_new, spp_boxes, normalized_boxes, categids, masksblob, instance_wts = get_blobs(
img, boxes.astype(np.float32), categids, masks)
#save blobs in private dict
self.myblobs['image'] = im_new.astype(np.float32)
self.myblobs['normalizedboxes'] = normalized_boxes.astype(np.float32)
self.myblobs['sppboxes'] = spp_boxes.astype(np.float32)
self.myblobs['categids'] = categids.astype(np.float32)
self.myblobs['labels'] = masksblob.astype(np.float32)
self.myblobs['instance_wts'] = instance_wts.astype(np.float32)
#and reshape
for i in range(len(top)):
top[i].reshape(*(self.myblobs[self.blob_names[i]].shape))
def reshape(self, bottom, top):
#sample a category
categid = np.random.choice(self.numclasses)
#sample an image for this category
imid = self.data_percateg[categid]['imids'][np.random.choice(len(self.data_percateg[categid]['imids']))]
img = cv2.imread(self._params.imgpath.format(self.names[imid]))
#get all possibilities for this category
start = self.data_percateg[categid]['im_end_index'][imid]+1
stop = self.data_percateg[categid]['im_end_index'][imid+1]
#pick a box
idx = np.random.choice(np.arange(start,stop+1), cfg.TRAIN_SAMPLES_PER_IMG)
boxid = self.data_percateg[categid]['boxids'][idx]
boxes = self.boxes[imid][boxid,:]-1
instid = self.data_percateg[categid]['instids'][idx]
#load the gt
[inst, categories] = sbd.load_gt(self.names[imid])
masks = np.zeros((idx.size, 1,inst.shape[0],inst.shape[1]))
for k in range(idx.size):
masks[k,0,:,:] = (inst==instid[k]+1).astype(np.float32)
categids = categid*np.ones(idx.size)
#get the blobs
im_new, spp_boxes, normalized_boxes, categids, masksblob, instance_wts = get_blobs(img, boxes.astype(np.float32), categids, masks)
#save blobs in private dict
self.myblobs['image']=im_new.astype(np.float32)
self.myblobs['normalizedboxes']=normalized_boxes.astype(np.float32)
self.myblobs['sppboxes']=spp_boxes.astype(np.float32)
self.myblobs['categids']=categids.astype(np.float32)
self.myblobs['labels']=masksblob.astype(np.float32)
self.myblobs['instance_wts']=instance_wts.astype(np.float32)
#and reshape
for i in range(len(top)):
top[i].reshape(*(self.myblobs[self.blob_names[i]].shape))
def get_all_outputs(net,names, dets, imgpath, sppath, regsppath,thresh=0.4,outpath=None, do_eval = True, eval_thresh = [0.5, 0.7]):
numcategs=dets['boxes'].size
if do_eval:
#we will accumulate the overlaps and the classes of the gt
all_ov=[]
gt = []
for j in range(numcategs):
all_ov.append([])
#a dictionary of times
times = {}
times['boxes']=0.
times['pred']=0.
times['sp']=0.
times['ov']=0.
times['total']=0.
for i in range(len(names)):
t1=time.time()
img = cv2.imread(imgpath.format(names[i]))
#get all boxes for this image
boxes_img = np.zeros((0,4))
cids_img = np.zeros((0,1))
for j in range(numcategs):
boxes_img = np.vstack((boxes_img, dets['boxes'][j][i]))
cids_img = np.vstack((cids_img, j*np.ones((dets['boxes'][j][i].shape[0],1))))
t2=time.time()
times['boxes']=times['boxes']+t2-t1
#get the predictions
output = get_hypercolumn_prediction(net, img, boxes_img.astype(np.float32), cids_img)
t3=time.time()
times['pred']=times['pred']+t3-t2
#project to sp
(sp, reg2sp) = sprep.read_sprep(sppath.format(names[i]), regsppath.format(names[i]))
newreg2sp_all = paste_output_sp(output, boxes_img.astype(np.float32)-1., sp.shape, sp)
newreg2sp_all = np.squeeze(newreg2sp_all)
newreg2sp_all = newreg2sp_all>=thresh
newreg2sp_all = newreg2sp_all.T
t4=time.time()
times['sp'] = times['sp']+t4-t3
#save if needed
if outpath is not None:
savemat(outpath.format(names[i]), {'output':output})
#evaluate
if do_eval:
inst, categories = sbd.load_gt(names[i])
ov = evaluation.compute_overlap_sprep(sp, newreg2sp_all, inst)
#separate according to categories
for j in range(numcategs):
all_ov[j].append(ov[:,cids_img.reshape(-1)==j])
#append categories
gt.append(np.squeeze(categories-1))
t5=time.time()
times['ov'] = times['ov']+t5-t4
if i % 100 == 0:
total = float(i+1)
print 'Doing : {:d}, get boxes:{:.2f} s, get pred:{:.2f} s, get sp:{:.2f} s, get ov:{:.2f} s'.format(i, times['boxes']/total,
times['pred']/total, times['sp']/total, times['ov']/total)
ap = [[] for _ in eval_thresh]
prec = [[] for _ in eval_thresh]
rec = [[] for _ in eval_thresh]
for i in range(numcategs):
print 'Evaluating :{:d}'.format(i)
for t,thr in enumerate(eval_thresh):
ap_, prec_, rec_ = evaluation.generalized_det_eval_simple(dets['scores'][i].tolist()[0:len(names)], all_ov[i], gt, i, thr)
ap[t].append(ap_)
prec[t].append(prec_)
rec[t].append(rec_)
return ap, prec, rec, all_ov, gt
