Fluid Faster RCNN implementation
Created by: Noplz
According to Detectron the ops need to be implemented for Faster RCNN:
| Detectron | Paddle Fluid |
|---|---|
| add_rpn_blobs |
generate_anchors_op iou_similarity_op fg_bg_target_assign_op argmax_op box_coder_op |
| generate_proposals | box_coder_op clip_and_filter_op multiclass_nms_op |
| generate_proposal_labels |
add_proposals_op add_gt_proposals_op target_assign_op |
Faster RCNN on Fluid
-
generate_anchors_op: generate anchors on the fly corresponding to configs
- Input:
- Attr:
- Output: anchor boxes with shape [H, W, M, 4]. H * W * M is total number of anchor box
-
fg_bg_target_assign_op:
- Input: Input(1): the anchor boxes with shape [M, 4] Input(2): the gt boxes with shape [N, 4]
- Attr:
- Output: anchor to gt matched mask with shape [M, 1]
-
box_coder_op:
- Input: Same
- Attr: Attr(1): add normalize(bool)
- Output: Same
-
clip_and_fliter_op: clip and filter proposals corresponding to input image
- Input: Input(1): decoded anchors with shape [M, 4]
- Attr: Attr(1): image width and height
- Output:
Output(1):cliped and filted proposals with shape [K, 1]
-
add_gt_proposals_op:
- Input: Input(1): the proposals with shape [M, 4] Input(2): the gt boxes with shape [N, 4]
- Attr:
- Output: the proposals with shape [M+N, 4]
-
add_proposals_op:
- Input:
- Attr:
- Output:
-
target_assign_op:
- Input:
- Attr:
- Output: