why need add offset？

[fenglifeng]

in my result，in detect function the bbox is already the coordinate of image include offset，so why need add offset again in to_coord function？

def to_coord(bbox, origin, space):
d = len(bbox.shape)
if d==1:
bbox = np.expand_dims(bbox,0)
space = space[::-1]
origin = origin[::-1]
center = (bbox[:,3:] + bbox[:,:3])/2*space + origin
center = center[:,::-1]
diam = l2norm((bbox[:,3:] - bbox[:,:3])*space)/3**0.5
coord = np.concatenate([center, diam],1)
if d==1:
coord = coord.flatten()
return coord

this is my code：
def detect(model, sample, batch_size=32):
candidates = []
for i, pixel_values in enumerate(torch.split(sample["pixel_values"].to(model.device), batch_size)):
img_shape = np.tile(np.array(pixel_values.shape[-3:]),2)
offsets = np.tile(sample["offsets"][i*batch_size:(i+1)*batch_size],2)
outputs = model(pixel_values=pixel_values)

    bbox = outputs.bbox.cpu().numpy()*img_shape+offsets
    coord = to_coord(bbox, sample["origin"], sample["space"])
    logits = outputs.logits.cpu().numpy()
    probabilities = sigmoid(logits)
    candidates.append(np.concatenate([coord, probabilities], 1))
    #candidates.append(np.concatenate([coord,logits],1))

candidates = np.concatenate(candidates,0)
#candidates = merge_cands(candidates)
# threshold cutoff
candidates = candidates[candidates[:,-1]>0.8]
return candidates

def to_coord(bbox, origin, space):
d = len(bbox.shape)
if d==1:
bbox = np.expand_dims(bbox,0)
space = space[::-1]
origin = origin[::-1]
center = (bbox[:,3:] + bbox[:,:3])/2
center = center[:,::-1]
diam = l2norm((bbox[:,3:] - bbox[:,:3])*space)/3**0.5
coord = np.concatenate([center, diam],1)
if d==1:
coord = coord.flatten()
return coord

def l2norm(x):
return np.sum(x**2,axis=-1,keepdims=True)**0.5

def sigmoid(x):
return 1 / (1 + np.exp(-x))