def get_groundtruth(self, index):
width, height = self.im_sizes[index, :]
# get object bounding boxes, labels and relations
obj_boxes = self.gt_boxes[index].copy()
obj_labels = self.gt_classes[index].copy()
obj_relation_triplets = self.relationships[index].copy()
if self.filter_duplicate_rels:
# Filter out dupes!
assert self.split == 'train'
old_size = obj_relation_triplets.shape[0]
all_rel_sets = defaultdict(list)
for (o0, o1, r) in obj_relation_triplets:
all_rel_sets[(o0, o1)].append(r)
obj_relation_triplets = [(k[0], k[1], np.random.choice(v))
for k, v in all_rel_sets.items()]
obj_relation_triplets = np.array(obj_relation_triplets)
obj_relations = np.zeros((obj_boxes.shape[0], obj_boxes.shape[0]))
for i in range(obj_relation_triplets.shape[0]):
subj_id = obj_relation_triplets[i][0]
obj_id = obj_relation_triplets[i][1]
pred = obj_relation_triplets[i][2]
obj_relations[subj_id, obj_id] = pred
target = BoxList(obj_boxes, (width, height), mode="xyxy")
target.add_field("labels", torch.from_numpy(obj_labels))
target.add_field("pred_labels", torch.from_numpy(obj_relations))
target.add_field("relation_labels",
torch.from_numpy(obj_relation_triplets))
target.add_field("difficult",
torch.from_numpy(obj_labels).clone().fill_(0))
return target
def im_detect_bbox_aug(model, images, device):
# Collect detections computed under different transformations
boxlists_ts = []
for _ in range(len(images)):
boxlists_ts.append([])
def add_preds_t(boxlists_t):
for i, boxlist_t in enumerate(boxlists_t):
if len(boxlists_ts[i]) == 0:
# The first one is identity transform, no need to resize the boxlist
boxlists_ts[i].append(boxlist_t)
else:
# Resize the boxlist as the first one
boxlists_ts[i].append(boxlist_t.resize(boxlists_ts[i][0].size))
# Compute detections for the original image (identity transform)
boxlists_i = im_detect_bbox(
model, images, cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MAX_SIZE_TEST, device
)
add_preds_t(boxlists_i)
# Perform detection on the horizontally flipped image
if cfg.TEST.BBOX_AUG.H_FLIP:
boxlists_hf = im_detect_bbox_hflip(
model, images, cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MAX_SIZE_TEST, device
)
add_preds_t(boxlists_hf)
# Compute detections at different scales
for scale in cfg.TEST.BBOX_AUG.SCALES:
max_size = cfg.TEST.BBOX_AUG.MAX_SIZE
boxlists_scl = im_detect_bbox_scale(
model, images, scale, max_size, device
)
add_preds_t(boxlists_scl)
if cfg.TEST.BBOX_AUG.SCALE_H_FLIP:
boxlists_scl_hf = im_detect_bbox_scale(
model, images, scale, max_size, device, hflip=True
)
add_preds_t(boxlists_scl_hf)
# Merge boxlists detected by different bbox aug params
boxlists = []
for i, boxlist_ts in enumerate(boxlists_ts):
bbox = torch.cat([boxlist_t.bbox for boxlist_t in boxlist_ts])
scores = torch.cat([boxlist_t.get_field('scores') for boxlist_t in boxlist_ts])
boxlist = BoxList(bbox, boxlist_ts[0].size, boxlist_ts[0].mode)
boxlist.add_field('scores', scores)
boxlists.append(boxlist)
# Apply NMS and limit the final detections
results = []
post_processor = make_roi_box_post_processor(cfg)
for boxlist in boxlists:
results.append(post_processor.filter_results(boxlist, cfg.MODEL.ROI_BOX_HEAD.NUM_CLASSES))
return results
def forward_for_single_feature_map(self, anchors, objectness,
box_regression):
"""
Arguments:
anchors: list[BoxList]
objectness: tensor of size N, A, H, W
box_regression: tensor of size N, A * 4, H, W
"""
device = objectness.device
N, A, H, W = objectness.shape
# put in the same format as anchors
objectness = permute_and_flatten(objectness, N, A, 1, H, W).view(N, -1)
objectness = objectness.sigmoid()
box_regression = permute_and_flatten(box_regression, N, A, 4, H, W)
num_anchors = A * H * W
pre_nms_top_n = min(self.pre_nms_top_n, num_anchors)
objectness, topk_idx = objectness.topk(pre_nms_top_n,
dim=1,
sorted=True)
batch_idx = torch.arange(N, device=device)[:, None]
box_regression = box_regression[batch_idx, topk_idx]
image_shapes = [box.size for box in anchors]
concat_anchors = torch.cat([a.bbox for a in anchors], dim=0)
concat_anchors = concat_anchors.reshape(N, -1, 4)[batch_idx, topk_idx]
proposals = self.box_coder.decode(box_regression.view(-1, 4),
concat_anchors.view(-1, 4))
proposals = proposals.view(N, -1, 4)
result = []
for proposal, score, im_shape in zip(proposals, objectness,
image_shapes):
boxlist = BoxList(proposal, im_shape, mode="xyxy")
boxlist.add_field("objectness", score)
boxlist = boxlist.clip_to_image(remove_empty=False)
boxlist = remove_small_boxes(boxlist, self.min_size)
boxlist = boxlist_nms(
boxlist,
self.nms_thresh,
max_proposals=self.post_nms_top_n,
score_field="objectness",
)
result.append(boxlist)
return result
def __getitem__(self, index):
"""
get dataset item
"""
# get image
ref = self.refer.Refs[self.ref_ids[index]]
img_id = ref["image_id"]
ann_id = ref["ann_id"]
img_path = os.path.join(self.refer.IMAGE_DIR, self.refer.Imgs[img_id]["file_name"])
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
#print("img original size", img.shape)
width, height = img.shape[0], img.shape[1]
#print("img after size", img.shape)
ref_expr = "\n".join([s["raw"] for s in ref["sentences"]])
# get object bounding boxes, labels and relations
#obj_boxes = [[34.79, 272.54, 106.72, 80.43]] # dummy target bbox
referent_box = [self.refer.Anns[ann_id]["bbox"]]
target_raw = BoxList(referent_box, (width, height), mode="xyxy")
if self.transforms is not None:
img, target = self.transforms(img, target_raw)
else:
img, target = img, target_raw
target.add_field("ref_sents", [s["raw"] for s in ref["sentences"]])
target.add_field("label", self.refer.Anns[ann_id]["category_id"])
#target.add_field("labels", torch.from_numpy(obj_labels))
#target.add_field("pred_labels", torch.from_numpy(obj_relations))
#target.add_field("relation_labels", torch.from_numpy(obj_relation_triplets))
target = target.clip_to_image(remove_empty=False)
info = {"img_id":img_id, "ann_id":ann_id, "ref_id": self.ref_ids[index], "ref_sents": [s["raw"] for s in ref["sentences"]]}
return img, target, index, info
def get_triplets_as_string(self, top_obj: BoxList,
top_pred: BoxPairList) -> List[str]:
"""
Given top detected objects and top predicted relationships, return
the triplets in human-readable form.
:param top_obj: BoxList containing top detected objects
:param top_pred: BoxPairList containing the top detected triplets
:return: List of triplets (in decreasing score order)
"""
# num_detected_objects
obj_indices = top_obj.get_field("labels")
# 100 x 2 (indices in obj_indices)
obj_pairs_indices = top_pred.get_field("idx_pairs")
# 100 (indices in GLOBAL relationship indices list)
rel_indices = top_pred.get_field("scores").max(1)[1]
# 100 x 3
top_triplets = torch.stack(
(obj_indices[obj_pairs_indices[:, 0]],
obj_indices[obj_pairs_indices[:, 1]], rel_indices), 1).tolist()
idx_to_obj = self.data_loader_test.dataset.ind_to_classes
idx_to_rel = self.data_loader_test.dataset.ind_to_predicates
# convert integers to labels
top_triplets_str = []
for t in top_triplets:
top_triplets_str.append(idx_to_obj[t[0]] + " " + idx_to_rel[t[2]] +
" " + idx_to_obj[t[1]])
return top_triplets_str
def __getitem__(self, index):
"""
get dataset item
"""
# get image
img_id = self.img_ids[index]
img_path = self.img_paths[index]
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
#print("img original size", img.shape)
width, height = img.shape[0], img.shape[1]
#print("img after size", img.shape)
ref_expr = "\n".join([])
# get object bounding boxes, labels and relations
referent_box = [[34.79, 272.54, 106.72, 80.43]] # dummy target bbox
#referent_box = [self.refer.Anns[ann_id]["bbox"]]
target_raw = BoxList(referent_box, (width, height), mode="xyxy")
if self.transforms is not None:
img, target = self.transforms(img, target_raw)
else:
img, target = img, target_raw
target.add_field("ref_sents", [])
target.add_field("label", None)
target = target.clip_to_image(remove_empty=False)
info = {"img_id": img_id}
return img, target, index, info
def __getitem__(self, index):
"""
get dataset item
"""
# get image
img = Image.fromarray(self._im_getter(index))
width, height = img.size
if self.NO_EVAL:
target_raw = BoxList([[0, 0, 0, 0]], (width, height), mode="xyxy")
img, target = self.transforms(img, target_raw)
return img, target, index
# get object bounding boxes, labels and relations
obj_boxes = self.gt_boxes[index].copy()
obj_labels = self.gt_classes[index].copy()
obj_relation_triplets = self.relationships[index].copy()
# TODO(cjrd) can we use this logic for inference?
if self.filter_duplicate_rels:
# Filter out dupes!
assert self.split == 'train'
old_size = obj_relation_triplets.shape[0]
all_rel_sets = defaultdict(list)
for (o0, o1, r) in obj_relation_triplets:
all_rel_sets[(o0, o1)].append(r)
obj_relation_triplets = [(k[0], k[1], np.random.choice(v))
for k, v in all_rel_sets.items()]
obj_relation_triplets = np.array(obj_relation_triplets)
obj_relations = np.zeros((obj_boxes.shape[0], obj_boxes.shape[0]))
for i in range(obj_relation_triplets.shape[0]):
subj_id = obj_relation_triplets[i][0]
obj_id = obj_relation_triplets[i][1]
pred = obj_relation_triplets[i][2]
obj_relations[subj_id, obj_id] = pred
target_raw = BoxList(obj_boxes, (width, height), mode="xyxy")
img, target = self.transforms(img, target_raw)
target.add_field("labels", torch.from_numpy(obj_labels))
target.add_field("pred_labels", torch.from_numpy(obj_relations))
target.add_field("relation_labels",
torch.from_numpy(obj_relation_triplets))
target = target.clip_to_image(remove_empty=False)
return img, target, index
def prepare_boxlist(self, boxes, scores, image_shape):
"""
Returns BoxList from `boxes` and adds probability scores information
as an extra field
`boxes` has shape (#detections, 4 * #classes), where each row represents
a list of predicted bounding boxes for each of the object classes in the
dataset (including the background class). The detections in each row
originate from the same object proposal.
`scores` has shape (#detection, #classes), where each row represents a list
of object detection confidence scores for each of the object classes in the
dataset (including the background class). `scores[i, j]`` corresponds to the
box at `boxes[i, j * 4:(j + 1) * 4]`.
"""
boxes = boxes.reshape(-1, 4)
scores = scores.reshape(-1)
boxlist = BoxList(boxes, image_shape, mode="xyxy")
boxlist.add_field("scores", scores)
return boxlist
def filter_results(self, boxlist, num_classes):
"""Returns bounding-box detection results by thresholding on scores and
applying non-maximum suppression (NMS).
"""
# unwrap the boxlist to avoid additional overhead.
# if we had multi-class NMS, we could perform this directly on the boxlist
boxes = boxlist.bbox.reshape(-1, num_classes * 4)
scores = boxlist.get_field("scores").reshape(-1, num_classes)
logits = BoxList.get_field("logits").reshape(-1, num_classes)
features = boxlist.get_field("features")
device = scores.device
result = []
# Apply threshold on detection probabilities and apply NMS
# Skip j = 0, because it's the background class
inds_all = scores > self.score_thresh
for j in range(1, num_classes):
inds = inds_all[:, j].nonzero().squeeze(1)
scores_j = scores[inds, j]
features_j = features[inds]
boxes_j = boxes[inds, j * 4:(j + 1) * 4]
boxlist_for_class = BoxList(boxes_j, boxlist.size, mode="xyxy")
boxlist_for_class.add_field("scores", scores_j)
boxlist_for_class.add_field("features", features_j)
boxlist_for_class = boxlist_nms(boxlist_for_class, self.nms)
num_labels = len(boxlist_for_class)
boxlist_for_class.add_field(
"labels",
torch.full((num_labels, ), j, dtype=torch.int64,
device=device))
result.append(boxlist_for_class)
result = cat_boxlist(result)
number_of_detections = len(result)
# Limit to max_per_image detections **over all classes**
if number_of_detections > self.detections_per_img > 0:
cls_scores = result.get_field("scores")
image_thresh, _ = torch.kthvalue(
cls_scores.cpu(),
number_of_detections - self.detections_per_img + 1)
keep = cls_scores >= image_thresh.item()
keep = torch.nonzero(keep).squeeze(1)
result = result[keep]
return result
def forward(self, image_list, feature_maps):
grid_sizes = [feature_map.shape[-2:] for feature_map in feature_maps]
anchors_over_all_feature_maps = self.grid_anchors(grid_sizes)
anchors = []
for i, (image_height,
image_width) in enumerate(image_list.image_sizes):
anchors_in_image = []
for anchors_per_feature_map in anchors_over_all_feature_maps:
boxlist = BoxList(anchors_per_feature_map,
(image_width, image_height),
mode="xyxy")
self.add_visibility_to(boxlist)
anchors_in_image.append(boxlist)
anchors.append(anchors_in_image)
return anchors
def select_over_all_levels(self, boxlists):
num_images = len(boxlists)
results = []
for i in range(num_images):
scores = boxlists[i].get_field("scores")
labels = boxlists[i].get_field("labels")
boxes = boxlists[i].bbox
boxlist = boxlists[i]
result = []
# skip the background
for j in range(1, self.num_classes):
inds = (labels == j).nonzero().view(-1)
scores_j = scores[inds]
boxes_j = boxes[inds, :].view(-1, 4)
boxlist_for_class = BoxList(boxes_j, boxlist.size, mode="xyxy")
boxlist_for_class.add_field("scores", scores_j)
boxlist_for_class = boxlist_nms(
boxlist_for_class, self.nms_thresh,
score_field="scores"
)
num_labels = len(boxlist_for_class)
boxlist_for_class.add_field(
"labels", torch.full((num_labels,), j,
dtype=torch.int64,
device=scores.device)
)
result.append(boxlist_for_class)
result = cat_boxlist(result)
number_of_detections = len(result)
# Limit to max_per_image detections **over all classes**
if number_of_detections > self.fpn_post_nms_top_n > 0:
cls_scores = result.get_field("scores")
image_thresh, _ = torch.kthvalue(
cls_scores.cpu(),
number_of_detections - self.fpn_post_nms_top_n + 1
)
keep = cls_scores >= image_thresh.item()
keep = torch.nonzero(keep).squeeze(1)
result = result[keep]
results.append(result)
return results
def forward_for_single_feature_map(
self, anchors, box_cls, box_regression):
"""
Arguments:
anchors: list[BoxList]
box_cls: tensor of size N, A * C, H, W
box_regression: tensor of size N, A * 4, H, W
"""
device = box_cls.device
N, _, H, W = box_cls.shape
A = box_regression.size(1) // 4
C = box_cls.size(1) // A
# put in the same format as anchors
box_cls = permute_and_flatten(box_cls, N, A, C, H, W)
box_cls = box_cls.sigmoid()
box_regression = permute_and_flatten(box_regression, N, A, 4, H, W)
box_regression = box_regression.reshape(N, -1, 4)
num_anchors = A * H * W
candidate_inds = box_cls > self.pre_nms_thresh
pre_nms_top_n = candidate_inds.view(N, -1).sum(1)
pre_nms_top_n = pre_nms_top_n.clamp(max=self.pre_nms_top_n)
results = []
for per_box_cls, per_box_regression, per_pre_nms_top_n, \
per_candidate_inds, per_anchors in zip(
box_cls,
box_regression,
pre_nms_top_n,
candidate_inds,
anchors):
# Sort and select TopN
# TODO most of this can be made out of the loop for
# all images.
# TODO:Yang: Not easy to do. Because the numbers of detections are
# different in each image. Therefore, this part needs to be done
# per image.
per_box_cls = per_box_cls[per_candidate_inds]
per_box_cls, top_k_indices = \
per_box_cls.topk(per_pre_nms_top_n, sorted=False)
per_candidate_nonzeros = \
per_candidate_inds.nonzero()[top_k_indices, :]
per_box_loc = per_candidate_nonzeros[:, 0]
per_class = per_candidate_nonzeros[:, 1]
per_class += 1
detections = self.box_coder.decode(
per_box_regression[per_box_loc, :].view(-1, 4),
per_anchors.bbox[per_box_loc, :].view(-1, 4)
)
boxlist = BoxList(detections, per_anchors.size, mode="xyxy")
boxlist.add_field("labels", per_class)
boxlist.add_field("scores", per_box_cls)
boxlist = boxlist.clip_to_image(remove_empty=False)
boxlist = remove_small_boxes(boxlist, self.min_size)
results.append(boxlist)
return results
def calc_detection_voc_prec_rec(gt_boxlists, pred_boxlists, iou_thresh=0.5):
"""Calculate precision and recall based on evaluation code of PASCAL VOC.
This function calculates precision and recall of
predicted bounding boxes obtained from a dataset which has :math:`N`
images.
The code is based on the evaluation code used in PASCAL VOC Challenge.
"""
n_pos = defaultdict(int)
score = defaultdict(list)
match = defaultdict(list)
for gt_boxlist, pred_boxlist in zip(gt_boxlists, pred_boxlists):
pred_bbox = pred_boxlist.bbox.numpy()
pred_label = pred_boxlist.get_field("labels").numpy()
pred_score = pred_boxlist.get_field("scores").numpy()
gt_bbox = gt_boxlist.bbox.numpy()
gt_label = gt_boxlist.get_field("labels").numpy()
gt_difficult = gt_boxlist.get_field("difficult").numpy()
for l in np.unique(np.concatenate((pred_label, gt_label)).astype(int)):
pred_mask_l = pred_label == l
pred_bbox_l = pred_bbox[pred_mask_l]
pred_score_l = pred_score[pred_mask_l]
# sort by score
order = pred_score_l.argsort()[::-1]
pred_bbox_l = pred_bbox_l[order]
pred_score_l = pred_score_l[order]
gt_mask_l = gt_label == l
gt_bbox_l = gt_bbox[gt_mask_l]
gt_difficult_l = gt_difficult[gt_mask_l]
n_pos[l] += np.logical_not(gt_difficult_l).sum()
score[l].extend(pred_score_l)
if len(pred_bbox_l) == 0:
continue
if len(gt_bbox_l) == 0:
match[l].extend((0, ) * pred_bbox_l.shape[0])
continue
# VOC evaluation follows integer typed bounding boxes.
pred_bbox_l = pred_bbox_l.copy()
pred_bbox_l[:, 2:] += 1
gt_bbox_l = gt_bbox_l.copy()
gt_bbox_l[:, 2:] += 1
iou = boxlist_iou(
BoxList(pred_bbox_l, gt_boxlist.size),
BoxList(gt_bbox_l, gt_boxlist.size),
).numpy()
gt_index = iou.argmax(axis=1)
# set -1 if there is no matching ground truth
gt_index[iou.max(axis=1) < iou_thresh] = -1
del iou
selec = np.zeros(gt_bbox_l.shape[0], dtype=bool)
for gt_idx in gt_index:
if gt_idx >= 0:
if gt_difficult_l[gt_idx]:
match[l].append(-1)
else:
if not selec[gt_idx]:
match[l].append(1)
else:
match[l].append(0)
selec[gt_idx] = True
else:
match[l].append(0)
n_fg_class = max(n_pos.keys()) + 1
prec = [None] * n_fg_class
rec = [None] * n_fg_class
for l in n_pos.keys():
score_l = np.array(score[l])
match_l = np.array(match[l], dtype=np.int8)
order = score_l.argsort()[::-1]
match_l = match_l[order]
tp = np.cumsum(match_l == 1)
fp = np.cumsum(match_l == 0)
# If an element of fp + tp is 0,
# the corresponding element of prec[l] is nan.
prec[l] = tp / (fp + tp)
# If n_pos[l] is 0, rec[l] is None.
if n_pos[l] > 0:
rec[l] = tp / n_pos[l]
return prec, rec
def evaluate_box_proposals(predictions,
dataset,
thresholds=None,
area="all",
limit=None):
"""Evaluate detection proposal recall metrics. This function is a much
faster alternative to the official COCO API recall evaluation code. However,
it produces slightly different results.
"""
# Record max overlap value for each gt box
# Return vector of overlap values
areas = {
"all": 0,
"small": 1,
"medium": 2,
"large": 3,
"96-128": 4,
"128-256": 5,
"256-512": 6,
"512-inf": 7,
}
area_ranges = [
[0**2, 1e5**2], # all
[0**2, 32**2], # small
[32**2, 96**2], # medium
[96**2, 1e5**2], # large
[96**2, 128**2], # 96-128
[128**2, 256**2], # 128-256
[256**2, 512**2], # 256-512
[512**2, 1e5**2],
] # 512-inf
assert area in areas, "Unknown area range: {}".format(area)
area_range = area_ranges[areas[area]]
gt_overlaps = []
num_pos = 0
for image_id, prediction in enumerate(predictions):
original_id = image_id # dataset.id_to_img_map[image_id]
img_info = dataset.get_img_info(image_id)
image_width = img_info["width"]
image_height = img_info["height"]
prediction = prediction.resize((image_width, image_height))
# sort predictions in descending order
# TODO maybe remove this and make it explicit in the documentation
inds = prediction.get_field("objectness").sort(descending=True)[1]
prediction = prediction[inds]
ann_ids = dataset.coco.getAnnIds(imgIds=original_id)
anno = dataset.coco.loadAnns(ann_ids)
gt_boxes = [obj["bbox"] for obj in anno if obj["iscrowd"] == 0]
gt_boxes = torch.as_tensor(gt_boxes).reshape(
-1, 4) # guard against no boxes
gt_boxes = BoxList(gt_boxes, (image_width, image_height),
mode="xywh").convert("xyxy")
gt_areas = torch.as_tensor(
[obj["area"] for obj in anno if obj["iscrowd"] == 0])
if len(gt_boxes) == 0:
continue
valid_gt_inds = (gt_areas >= area_range[0]) & (gt_areas <=
area_range[1])
gt_boxes = gt_boxes[valid_gt_inds]
num_pos += len(gt_boxes)
if len(gt_boxes) == 0:
continue
if len(prediction) == 0:
continue
if limit is not None and len(prediction) > limit:
prediction = prediction[:limit]
overlaps = boxlist_iou(prediction, gt_boxes)
_gt_overlaps = torch.zeros(len(gt_boxes))
for j in range(min(len(prediction), len(gt_boxes))):
# find which proposal box maximally covers each gt box
# and get the iou amount of coverage for each gt box
max_overlaps, argmax_overlaps = overlaps.max(dim=0)
# find which gt box is 'best' covered (i.e. 'best' = most iou)
gt_ovr, gt_ind = max_overlaps.max(dim=0)
assert gt_ovr >= 0
# find the proposal box that covers the best covered gt box
box_ind = argmax_overlaps[gt_ind]
# record the iou coverage of this gt box
_gt_overlaps[j] = overlaps[box_ind, gt_ind]
assert _gt_overlaps[j] == gt_ovr
# mark the proposal box and the gt box as used
overlaps[box_ind, :] = -1
overlaps[:, gt_ind] = -1
# append recorded iou coverage level
gt_overlaps.append(_gt_overlaps)
gt_overlaps = torch.cat(gt_overlaps, dim=0)
gt_overlaps, _ = torch.sort(gt_overlaps)
if thresholds is None:
step = 0.05
thresholds = torch.arange(0.5, 0.95 + 1e-5, step, dtype=torch.float32)
recalls = torch.zeros_like(thresholds)
# compute recall for each iou threshold
for i, t in enumerate(thresholds):
recalls[i] = (gt_overlaps >= t).float().sum() / float(num_pos)
# ar = 2 * np.trapz(recalls, thresholds)
ar = recalls.mean()
return {
"ar": ar,
"recalls": recalls,
"thresholds": thresholds,
"gt_overlaps": gt_overlaps,
"num_pos": num_pos,
}
def forward(self, x, boxes, skip_nms=False):
"""
Arguments:
x (tuple[tensor, tensor]): x contains the class logits
and the box_regression from the model.
boxes (list[BoxList]): bounding boxes that are used as
reference, one for ech image
Returns:
results (list[BoxList]): one BoxList for each image, containing
the extra fields labels and scores
"""
class_logit, box_regression = x
class_prob = F.softmax(class_logit, -1)
# TODO think about a representation of batch of boxes
image_shapes = [box.size for box in boxes]
boxes_per_image = [len(box) for box in boxes]
features = [box.get_field("features") for box in boxes]
concat_boxes = torch.cat([a.bbox for a in boxes], dim=0)
if not skip_nms:
if self.cls_agnostic_bbox_reg:
box_regression = box_regression[:, -4:]
proposals = self.box_coder.decode(
box_regression.view(sum(boxes_per_image), -1), concat_boxes)
if self.cls_agnostic_bbox_reg:
proposals = proposals.repeat(1, class_prob.shape[1])
proposals = proposals.split(boxes_per_image, dim=0)
else:
proposals = concat_boxes.split(boxes_per_image, dim=0)
num_classes = class_prob.shape[1]
class_prob = class_prob.split(boxes_per_image, dim=0)
class_logit = class_logit.split(boxes_per_image, dim=0)
results = []
idx = 0
for prob, logit, boxes_per_img, features_per_img, image_shape in zip(
class_prob, class_logit, proposals, features, image_shapes):
if not self.bbox_aug_enabled and not skip_nms: # If bbox aug is enabled, we will do it later
boxlist = self.prepare_boxlist(boxes_per_img, features_per_img,
prob, logit, image_shape)
boxlist = boxlist.clip_to_image(remove_empty=False)
if not self.relation_on:
boxlist_filtered = self.filter_results(
boxlist, num_classes)
else:
# boxlist_pre = self.filter_results(boxlist, num_classes)
boxlist_filtered = self.filter_results_nm(
boxlist, num_classes)
# to enforce minimum number of detections per image
# we will do a binary search on the confidence threshold
score_thresh = 0.05
while len(boxlist_filtered) < self.min_detections_per_img:
score_thresh /= 2.0
print(("\nNumber of proposals {} is too small, "
"retrying filter_results with score thresh"
" = {}").format(len(boxlist_filtered),
score_thresh))
boxlist_filtered = self.filter_results_nm(
boxlist, num_classes, thresh=score_thresh)
else:
boxlist = BoxList(boxes_per_img, image_shape, mode="xyxy")
boxlist.add_field("scores", prob[:, 1:].max(1)[0])
boxlist.add_field("logits", logit)
boxlist.add_field("features", features_per_img)
boxlist.add_field("labels", boxes[idx].get_field("labels"))
boxlist.add_field("regression_targets",
boxes[idx].bbox.clone().fill_(0.0))
boxlist_filtered = boxlist
idx += 1
if len(boxlist) == 0:
raise ValueError("boxlist shoud not be empty!")
results.append(boxlist_filtered)
return results
def filter_results_nm(self, boxlist, num_classes, thresh=0.05):
"""Returns bounding-box detection results by thresholding on scores and
applying non-maximum suppression (NMS). Similar to Neural-Motif Network
"""
# unwrap the boxlist to avoid additional overhead.
# if we had multi-class NMS, we could perform this directly on the boxlist
boxes = boxlist.bbox.reshape(-1, num_classes * 4)
scores = boxlist.get_field("scores").reshape(-1, num_classes)
logits = boxlist.get_field("logits").reshape(-1, num_classes)
features = boxlist.get_field("features")
valid_cls = (scores[:, 1:].max(0)[0] > thresh).nonzero() + 1
nms_mask = scores.clone()
nms_mask.zero_()
device = scores.device
result = []
# Apply threshold on detection probabilities and apply NMS
# Skip j = 0, because it's the background class
inds_all = scores > self.score_thresh
for j in valid_cls.view(-1).cpu():
scores_j = scores[:, j]
boxes_j = boxes[:, j * 4:(j + 1) * 4]
boxlist_for_class = BoxList(boxes_j, boxlist.size, mode="xyxy")
boxlist_for_class.add_field("scores", scores_j)
boxlist_for_class.add_field(
"idxs",
torch.arange(0, scores.shape[0]).long())
# boxlist_for_class = boxlist_nms(
# boxlist_for_class, self.nms
# )
boxlist_for_class = boxlist_nms(boxlist_for_class, 0.3)
nms_mask[:, j][boxlist_for_class.get_field("idxs")] = 1
num_labels = len(boxlist_for_class)
boxlist_for_class.add_field(
"labels",
torch.full((num_labels, ), j, dtype=torch.int64,
device=device))
result.append(boxlist_for_class)
dists_all = nms_mask * scores
# filter duplicate boxes
scores_pre, labels_pre = dists_all.max(1)
inds_all = scores_pre.nonzero()
assert inds_all.dim() != 0
inds_all = inds_all.squeeze(1)
labels_all = labels_pre[inds_all]
scores_all = scores_pre[inds_all]
features_all = features[inds_all]
logits_all = logits[inds_all]
box_inds_all = inds_all * scores.shape[1] + labels_all
result = BoxList(boxlist.bbox.view(-1, 4)[box_inds_all],
boxlist.size,
mode="xyxy")
result.add_field("labels", labels_all)
result.add_field("scores", scores_all)
result.add_field("logits", logits_all)
result.add_field("features", features_all)
number_of_detections = len(result)
vs, idx = torch.sort(scores_all, dim=0, descending=True)
idx = idx[vs > thresh]
if self.detections_per_img < idx.size(0):
idx = idx[:self.detections_per_img]
result = result[idx]
return result
v = v.draw_instance_predictions(pred)
showarray(v.get_image()[:, :, ::-1], save_dir + "pred_%i.jpg" % i)
#print('instances:\n', instances)
#print()
#print('boxes:\n', instances.pred_boxes)
#print()
#print('Shape of features:\n', features.shape)
#pred_class_logits, pred_attr_logits, pred_proposal_deltas = predictor.model.roi_heads.box_predictor(features)
#pred_class_probs = torch.nn.functional.softmax(pred_class_logits, -1)[:, :-1]
#max_probs, max_classes = pred_class_probs.max(-1)
#print("%d objects are different, it is because the classes-aware NMS process" % (NUM_OBJECTS - torch.eq(instances.pred_classes, max_classes).sum().item()))
#print("The total difference of score is %0.4f" % (instances.scores - max_probs).abs().sum().item())
boxes = instances.pred_boxes.tensor
boxlist = BoxList(boxes.cpu(), (image_w, image_h), mode="xyxy")
boxlist.add_field("scores", instances.scores.cpu())
boxlist.add_field("labels", instances.pred_classes.cpu())
boxlist.add_field("attr_logits", instances.attr_logits.cpu())
boxlist.add_field("cls_logits", instances.cls_logits.cpu())
data.append(boxlist)
#new_entry["boxes"] = instances.pred_boxes.tensor.cpu().numpy()
#new_entry["box_scores"] = instances.scores.cpu().numpy()
#new_entry["pred_classes"] = instances.pred_classes.cpu().numpy()
#new_entry["attr_logits"] = instances.attr_logits.cpu().numpy()
#new_entry["cls_logits"] = instances.cls_logits.cpu().numpy()
#np_dict = np.array(list(new_entry.items()))
#np.savetxt("./data/%i.npy" % i, np_dict)
#dd.io.save("./data/%i.h5" % i, new_entry, compression="default")
#data.append(new_entry)
#with open("./data/%i.json" % i, "w") as f: