Open Dantju opened 4 years ago
@Dantju We will release it but we do not have an ETA now. Please stay tuned.
How long will it take?
@tianzhi0549 I train the e2e_keypoint_rcnn_R_50_FPN_1x.yaml,but get error while test,It's caused by direct pose not in code?
@Dantju Sure. It's the config for Mask R-CNN based keypoint detection.
@tianzhi0549 so how can i change it to detect person and detect keypoint ,and how long will direct pose be released
@tianzhi0549 We do not have an estimated date for DirectPose release. You can try to change the number of outputs of FCOS to adapt it to keypoint detection.
@tianzhi0549 how long will it take to release directpose ,several month,several weeks or will not release,I find it‘s hard to modify code
@Dantju we will release our code if our paper gets accepted. It might take a few months.
@tianzhi0549 I said this paper has been accepted by CVPR
Comments: | 12 pages |
---|---|
Subjects: | Computer Vision and Pattern Recognition (cs.CV) |
Cite as: | arXiv:1911.07451 [cs.CV] |
(or arXiv:1911.07451v2 [cs.CV] for this version) |
@Dantju it was not accepted by CVPR.
@tianzhi0549 so sad,can u release part code,may be KPAlign is important,can u release code without KPAlign
@Dantju Sorry, we can't. But please feel free to ask your questions if you implement it by yourself.
@tianzhi0549 ,I add keypoint in fcos_head.py as follow:
def forward(self, x):
logits = []
bbox_reg = []
keypoint = []
centerness = []
centerness_kps = []
for l, feature in enumerate(x):
# print(feature.shape)
# print("###############fcos head##############")
cls_tower = self.cls_tower(feature)
logits.append(self.cls_logits(cls_tower))
box_tower = self.bbox_tower(feature)
keypoint_tower = self.keypoint_tower(feature)
centerness.append(self.centerness(box_tower))
centerness_kps.append(self.centerness_kps(keypoint_tower))
bbox_reg.append(torch.exp(self.scales[l](
self.bbox_pred(box_tower)
)))
keypoint.append(torch.exp(self.scales[l](
self.keypoint_pred(keypoint_tower)
)))
return logits, bbox_reg, centerness, keypoint
and in fcos.py: def _forward_train(self, locations, box_cls, box_regression, centerness, keypoint,targets, image_sizes): loss_box_cls, loss_box_reg, loss_centerness = self.loss_evaluator( locations, box_cls, box_regression, centerness, targets )
# proposals = []
boxes = self.box_selector_train(
locations, box_cls, box_regression,
centerness, image_sizes, targets
)
# print("###########fcos#######################")
# print(np.array(boxes).shape)
# print("############fcos#######################")
with torch.no_grad():
proposals = self.kps_loss_evaluator.subsample(boxes, targets)
# print("##################################")
# print(np.array(keypoint).shape)
# print("###################################")
loss_kp = self.kps_loss_evaluator(proposals, keypoint)
# kps = self.kps_post_processor(keypoint,proposals)
losses = {
"loss_cls": loss_box_cls,
"loss_reg": loss_box_reg,
"loss_centerness": loss_centerness,
"loss_keypoint": loss_kp
}
get error in modeling\roi_heads\keypoint_head\loss.py,say keypoint_logits is a list N, K, H, W = keypoint_logits.shape keypoint_logits = keypoint_logits.view(N K, H W)
@tianzhi0549 how to prepare_kps_target as (labels,reg_targets =self.prepare_targets(locations,targets))
@tianzhi0549 I want to know how to prepare gt for keypoints train
@tianzhi0549 what does locations & object_sizes_of_interest means
@tianzhi0549 I find u havn't do any preprocess,so when I make kps gt heatmap ,how to get kps in 5 level feature map,kps gt div (ori image size/feature map size)?
@tianzhi0549 pts is the keypoints gt,I get featuremap gt by using this,is it right
@tianzhi0549 when in level five,the stride is 128 ,use this may get negative keypoints, how to deal with this ?can I just pts = pts /scale_level[level] to get gt in feature map level?
@tianzhi0549 the point for classification is the center of gt bbox, right?
@tianzhi0549 how long direct pose will be released?
when the direct pose will be released?