def test_imrenormalize():
from mmtrack.core import imrenormalize
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True)
new_img_norm_cfg = dict(mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True)
img = np.random.randn(128, 256, 3).astype(np.float32)
new_img = imrenormalize(img, img_norm_cfg, new_img_norm_cfg)
assert isinstance(new_img, np.ndarray)
assert new_img.shape == (128, 256, 3)
assert np.allclose(img, new_img, atol=1e-6)
img = torch.randn(1, 3, 128, 256, dtype=torch.float)
new_img = imrenormalize(img, img_norm_cfg, new_img_norm_cfg)
assert isinstance(new_img, torch.Tensor)
assert new_img.shape == (1, 3, 128, 256)
assert np.allclose(img, new_img, atol=1e-6)
def track(self,
img,
img_metas,
model,
bboxes,
labels,
frame_id,
rescale=False,
**kwargs):
"""Tracking forward function.
Args:
img (Tensor): of shape (N, C, H, W) encoding input images.
Typically these should be mean centered and std scaled.
img_metas (list[dict]): list of image info dict where each dict
has: 'img_shape', 'scale_factor', 'flip', and may also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
model (nn.Module): MOT model.
bboxes (Tensor): of shape (N, 5).
labels (Tensor): of shape (N, ).
frame_id (int): The id of current frame, 0-index.
rescale (bool, optional): If True, the bounding boxes should be
rescaled to fit the original scale of the image. Defaults to
False.
Returns:
tuple: Tracking results.
"""
if not hasattr(self, 'kf'):
self.kf = model.motion
if self.with_reid:
if self.reid.get('img_norm_cfg', False):
reid_img = imrenormalize(img, img_metas[0]['img_norm_cfg'],
self.reid['img_norm_cfg'])
else:
reid_img = img.clone()
valid_inds = bboxes[:, -1] > self.obj_score_thr
bboxes = bboxes[valid_inds]
labels = labels[valid_inds]
if self.empty or bboxes.size(0) == 0:
num_new_tracks = bboxes.size(0)
ids = torch.arange(self.num_tracks,
self.num_tracks + num_new_tracks,
dtype=torch.long)
self.num_tracks += num_new_tracks
if self.with_reid:
embeds = model.reid.simple_test(
self.crop_imgs(reid_img, img_metas, bboxes[:, :4].clone(),
rescale))
else:
ids = torch.full((bboxes.size(0), ), -1, dtype=torch.long)
# motion
if model.with_motion:
self.tracks, costs = model.motion.track(
self.tracks, bbox_xyxy_to_cxcyah(bboxes))
active_ids = self.confirmed_ids
if self.with_reid:
embeds = model.reid.simple_test(
self.crop_imgs(reid_img, img_metas, bboxes[:, :4].clone(),
rescale))
# reid
if len(active_ids) > 0:
track_embeds = self.get('embeds',
active_ids,
self.reid.get('num_samples', None),
behavior='mean')
reid_dists = torch.cdist(track_embeds,
embeds).cpu().numpy()
valid_inds = [list(self.ids).index(_) for _ in active_ids]
reid_dists[~np.isfinite(costs[valid_inds, :])] = np.nan
row, col = linear_sum_assignment(reid_dists)
for r, c in zip(row, col):
dist = reid_dists[r, c]
if not np.isfinite(dist):
continue
if dist <= self.reid['match_score_thr']:
ids[c] = active_ids[r]
active_ids = [
id for id in self.ids if id not in ids
and self.tracks[id].frame_ids[-1] == frame_id - 1
]
if len(active_ids) > 0:
active_dets = torch.nonzero(ids == -1).squeeze(1)
track_bboxes = self.get('bboxes', active_ids)
ious = bbox_overlaps(
track_bboxes, bboxes[active_dets][:, :-1]).cpu().numpy()
dists = 1 - ious
row, col = linear_sum_assignment(dists)
for r, c in zip(row, col):
dist = dists[r, c]
if dist < 1 - self.match_iou_thr:
ids[active_dets[c]] = active_ids[r]
new_track_inds = ids == -1
ids[new_track_inds] = torch.arange(self.num_tracks,
self.num_tracks +
new_track_inds.sum(),
dtype=torch.long)
self.num_tracks += new_track_inds.sum()
self.update(ids=ids,
bboxes=bboxes[:, :4],
scores=bboxes[:, -1],
labels=labels,
embeds=embeds if self.with_reid else None,
frame_ids=frame_id)
return bboxes, labels, ids
def track(self,
img,
img_metas,
model,
feats,
bboxes,
labels,
frame_id,
rescale=False,
**kwargs):
"""Tracking forward function.
Args:
img (Tensor): of shape (N, C, H, W) encoding input images.
Typically these should be mean centered and std scaled.
img_metas (list[dict]): list of image info dict where each dict
has: 'img_shape', 'scale_factor', 'flip', and may also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
model (nn.Module): MOT model.
feats (tuple): Backbone features of the input image.
bboxes (Tensor): of shape (N, 5).
labels (Tensor): of shape (N, ).
frame_id (int): The id of current frame, 0-index.
rescale (bool, optional): If True, the bounding boxes should be
rescaled to fit the original scale of the image. Defaults to
False.
Returns:
tuple: Tracking results.
"""
if self.with_reid:
if self.reid.get('img_norm_cfg', False):
reid_img = imrenormalize(img, img_metas[0]['img_norm_cfg'],
self.reid['img_norm_cfg'])
else:
reid_img = img.clone()
valid_inds = bboxes[:, -1] > self.obj_score_thr
bboxes = bboxes[valid_inds]
labels = labels[valid_inds]
if self.empty:
num_new_tracks = bboxes.size(0)
ids = torch.arange(self.num_tracks,
self.num_tracks + num_new_tracks,
dtype=torch.long)
self.num_tracks += num_new_tracks
if self.with_reid:
embeds = model.reid.simple_test(
self.crop_imgs(reid_img, img_metas, bboxes[:, :4].clone(),
rescale))
else:
# motion
if model.with_cmc:
if model.with_linear_motion:
num_samples = model.linear_motion.num_samples
else:
num_samples = 1
self.tracks = model.cmc.track(self.last_img, img, self.tracks,
num_samples, frame_id)
if model.with_linear_motion:
self.tracks = model.linear_motion.track(self.tracks, frame_id)
# propagate tracks
prop_bboxes, prop_labels, prop_ids = self.regress_tracks(
feats, img_metas, model.detector, frame_id, rescale)
# filter bboxes with propagated tracks
ious = bbox_overlaps(bboxes[:, :4], prop_bboxes[:, :4])
valid_inds = (ious < self.regression['match_iou_thr']).all(dim=1)
bboxes = bboxes[valid_inds]
labels = labels[valid_inds]
ids = torch.full((bboxes.size(0), ), -1, dtype=torch.long)
if self.with_reid:
prop_embeds = model.reid.simple_test(
self.crop_imgs(reid_img, img_metas,
prop_bboxes[:, :4].clone(), rescale))
if bboxes.size(0) > 0:
embeds = model.reid.simple_test(
self.crop_imgs(reid_img, img_metas,
bboxes[:, :4].clone(), rescale))
else:
embeds = prop_embeds.new_zeros((0, prop_embeds.size(1)))
# reid
active_ids = [int(_) for _ in self.ids if _ not in prop_ids]
if len(active_ids) > 0 and bboxes.size(0) > 0:
track_embeds = self.get('embeds',
active_ids,
self.reid.get('num_samples', None),
behavior='mean')
reid_dists = torch.cdist(track_embeds,
embeds).cpu().numpy()
track_bboxes = self.get('bboxes', active_ids)
ious = bbox_overlaps(track_bboxes,
bboxes[:, :4]).cpu().numpy()
iou_masks = ious < self.reid['match_iou_thr']
reid_dists[iou_masks] = 1e6
row, col = linear_sum_assignment(reid_dists)
for r, c in zip(row, col):
dist = reid_dists[r, c]
if dist <= self.reid['match_score_thr']:
ids[c] = active_ids[r]
new_track_inds = ids == -1
ids[new_track_inds] = torch.arange(self.num_tracks,
self.num_tracks +
new_track_inds.sum(),
dtype=torch.long)
self.num_tracks += new_track_inds.sum()
if bboxes.shape[1] == 4:
bboxes = bboxes.new_zeros((0, 5))
if prop_bboxes.shape[1] == 4:
prop_bboxes = prop_bboxes.new_zeros((0, 5))
bboxes = torch.cat((prop_bboxes, bboxes), dim=0)
labels = torch.cat((prop_labels, labels), dim=0)
ids = torch.cat((prop_ids, ids), dim=0)
if self.with_reid:
embeds = torch.cat((prop_embeds, embeds), dim=0)
self.update(ids=ids,
bboxes=bboxes[:, :4],
scores=bboxes[:, -1],
labels=labels,
embeds=embeds if self.with_reid else None,
frame_ids=frame_id)
self.last_img = img
return bboxes, labels, ids
