[Feature] Support centernet dev1.x

docs/en/03-benchmark/benchmark.md

@@ -725,6 +725,20 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](../
     <td align="center">37.4</td>
     <td align="center">-</td>
   </tr>
+  <tr>
+    <td align="center"><a href="https://github.com/open-mmlab/mmdetection/tree/v3.0.0rc2/configs/centernet/centernet_r18_8xb16-crop512-140e_coco.py">CenterNet</a></td>
+    <td align="center">Object Detection</td>
+    <td align="center">COCO2017</td>
+    <td align="center">box AP</td>
+    <td align="center">25.9</td>
+    <td align="center">-</td>
+    <td align="center">-</td>
+    <td align="center">26.0</td>
+    <td align="center">-</td>
+    <td align="center">-</td>
+    <td align="center">-</td>
+    <td align="center">-</td>
+  </tr>
   <tr>
     <td align="center"><a href="https://github.com/open-mmlab/mmdetection/tree/master/configs/yolox/yolox_s_8x8_300e_coco.py">YOLOX</a></td>
     <td align="center">Object Detection</td>

docs/zh_cn/03-benchmark/benchmark.md

@@ -701,6 +701,20 @@ GPU: ncnn, TensorRT, PPLNN
     <td align="center">37.4</td>
     <td align="center">-</td>
   </tr>
+  <tr>
+    <td align="center"><a href="https://github.com/open-mmlab/mmdetection/tree/v3.0.0rc2/configs/centernet/centernet_r18_8xb16-crop512-140e_coco.py">CenterNet</a></td>
+    <td align="center">Object Detection</td>
+    <td align="center">COCO2017</td>
+    <td align="center">box AP</td>
+    <td align="center">25.9</td>
+    <td align="center">-</td>
+    <td align="center">-</td>
+    <td align="center">26.0</td>
+    <td align="center">-</td>
+    <td align="center">-</td>
+    <td align="center">-</td>
+    <td align="center">-</td>
+  </tr>
   <tr>
     <td align="center"><a href="https://github.com/open-mmlab/mmdetection/tree/master/configs/yolox/yolox_s_8x8_300e_coco.py">YOLOX</a></td>
     <td align="center">Object Detection</td>

mmdeploy/codebase/mmdet/deploy/object_detection_model.py

@@ -187,10 +187,18 @@ def forward(self,
 
             bboxes = dets[:, :4]
             scores = dets[:, 4]
-
             # perform rescale
-            if rescale and 'scale_factor' in img_metas[i]:
-                scale_factor = img_metas[i]['scale_factor']
+            if rescale:
+                if 'scale_factor' in img_metas[i]:
+                    scale_factor = img_metas[i]['scale_factor']
+                else:
+                    # Some models like CenterNet use rescale with out scale_factor
+                    # We multiply the feat_size inside the model, and divide the
+                    # image shape outside the model to adapt dynamic shape.
+                    scale_factor = np.array([
+                        1.0 / img_metas[i]['batch_input_shape'][1],
+                        1.0 / img_metas[i]['batch_input_shape'][0]
+                    ])
                 if isinstance(scale_factor, (list, tuple, np.ndarray)):
                     if len(scale_factor) == 2:
                         scale_factor = np.array(scale_factor)
@@ -200,12 +208,18 @@ def forward(self,
                 scale_factor = torch.from_numpy(scale_factor).to(dets)
                 bboxes /= scale_factor
 
-            if 'pad_param' in img_metas[i]:
+            if 'pad_param' in img_metas[i] or 'border' in img_metas[i]:
+                # Most of models in mmdetection 3.x use `pad_param`, but some
+                # models like CenterNet uses `border`.
                 # offset pixel of the top-left corners between original image
                 # and padded/enlarged image, 'pad_param' is used when exporting
                 # CornerNet and CentripetalNet to onnx
-                x_off = img_metas[i]['pad_param'][2]
-                y_off = img_metas[i]['pad_param'][0]
+                if 'pad_param' in img_metas[i]:
+                    x_off = img_metas[i]['pad_param'][2]
+                    y_off = img_metas[i]['pad_param'][0]
+                else:
+                    x_off = img_metas[i]['border'][2]
+                    y_off = img_metas[i]['border'][0]
                 bboxes[:, ::2] -= x_off
                 bboxes[:, 1::2] -= y_off
                 bboxes *= (bboxes > 0)

mmdeploy/codebase/mmdet/models/dense_heads/__init__.py

@@ -2,7 +2,8 @@
 from . import detr_head
 from .base_dense_head import (base_dense_head__predict_by_feat,
                               base_dense_head__predict_by_feat__ncnn)
-from .centernet_head import centernet_head__predict_by_feat__default
+from .centernet_head import (centernet_head__decode_heatmap__default,
+                             centernet_head__predict_by_feat__default)
 from .fovea_head import fovea_head__predict_by_feat
 from .gfl_head import gfl_head__predict_by_feat
 from .reppoints_head import reppoints_head__predict_by_feat
@@ -20,5 +21,6 @@
     'fovea_head__predict_by_feat', 'base_dense_head__predict_by_feat__ncnn',
     'yolox_head__predict_by_feat__ncnn', 'gfl_head__predict_by_feat',
     'reppoints_head__predict_by_feat', 'detr_head',
-    'centernet_head__predict_by_feat__default', 'rtmdet_head__predict_by_feat'
+    'centernet_head__predict_by_feat__default', 'rtmdet_head__predict_by_feat',
+    'centernet_head__decode_heatmap__default'
 ]

mmdeploy/codebase/mmdet/models/dense_heads/centernet_head.py

@@ -1,9 +1,12 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-from typing import List
+from typing import List, Tuple
 
 from torch import Tensor
+import torch
 
 from mmdeploy.core import FUNCTION_REWRITER
+from mmdet.models.utils import (get_local_maximum, get_topk_from_heatmap,
+                                transpose_and_gather_feat)
 
 
 @FUNCTION_REWRITER.register_rewriter(
@@ -37,14 +40,52 @@ def centernet_head__predict_by_feat__default(
         kernel=self.test_cfg.local_maximum_kernel)
     det_bboxes = batch_det_bboxes.reshape([batch_size, -1, 5])
     det_labels = batch_labels.reshape(batch_size, -1)
+    '''
     batch_border = det_bboxes.new_tensor(
         batch_img_metas[0]['border'])[..., [2, 0, 2, 0]]
     det_bboxes[..., :4] -= batch_border
+    print(f'debugging scale: {batch_img_metas[0]}')
     if rescale and 'scale_factor' in batch_img_metas[0]:
         det_bboxes[..., :4] /= det_bboxes.new_tensor(
             batch_img_metas[0]['scale_factor']).repeat((1, 2))
-
+    '''
     if with_nms:
         det_bboxes, det_labels = self._bboxes_nms(det_bboxes, det_labels,
                                                   self.test_cfg)
     return det_bboxes, det_labels
+
+
+@FUNCTION_REWRITER.register_rewriter(
+    'mmdet.models.dense_heads.centernet_head.CenterNetHead._decode_heatmap')
+def centernet_head__decode_heatmap__default(
+        ctx,
+        self,
+        center_heatmap_pred: Tensor,
+        wh_pred: Tensor,
+        offset_pred: Tensor,
+        img_shape: tuple,
+        k: int = 100,
+        kernel: int = 3) -> Tuple[Tensor, Tensor]:
+    """Rewrite `_decode_heatmap` of `CenterNetHead` for default backend."""
+
+    # Rewrite this function to move the img_shape calculation outside the
+    # model for dynamic shape deployment.
+    height, width = center_heatmap_pred.shape[2:]
+    center_heatmap_pred = get_local_maximum(center_heatmap_pred, kernel=kernel)
+
+    *batch_dets, topk_ys, topk_xs = get_topk_from_heatmap(
+        center_heatmap_pred, k=k)
+    batch_scores, batch_index, batch_topk_labels = batch_dets
+
+    wh = transpose_and_gather_feat(wh_pred, batch_index)
+    offset = transpose_and_gather_feat(offset_pred, batch_index)
+    topk_xs = topk_xs + offset[..., 0]
+    topk_ys = topk_ys + offset[..., 1]
+    tl_x = (topk_xs - wh[..., 0] / 2) / width
+    tl_y = (topk_ys - wh[..., 1] / 2) / height
+    br_x = (topk_xs + wh[..., 0] / 2) / width
+    br_y = (topk_ys + wh[..., 1] / 2) / height
+
+    batch_bboxes = torch.stack([tl_x, tl_y, br_x, br_y], dim=2)
+    batch_bboxes = torch.cat((batch_bboxes, batch_scores[..., None]), dim=-1)
+    return batch_bboxes, batch_topk_labels