fix label error, bbox error

Author: Leo24-m

.gitignore

@@ -127,3 +127,5 @@ work_dirs/
 
 *.png
 *.jpg
+
+/work_dirs/

custom_configs/DELIVER/deliver_dataset.py

@@ -8,9 +8,9 @@
 
 # Dataset basic info
 dataset_type = 'DELIVERDetectionDataset'
-data_root = '/media/jemo/HDD1/Workspace/dset/DELIVER/'  # Added trailing slash
+data_root = '/SSDb/jemo_maeng/dset/DELIVER/'  # Added trailing slash
 backend_args = None
-classes = ('Vehicle', 'Human')
+classes = ('Human', 'Vehicle')
 
 # Data preprocessor
 data_preprocessor = dict(
@@ -44,7 +44,11 @@
         prob=0.5,
         bbox_format='xywh'  # 🔥 xywh format 명시
     ),
-    dict(type='PackDELIVERDetInputs')
+    # dict(type='PackDELIVERDetInputs')
+    dict(
+    type='PackDELIVERDetInputs',
+    meta_keys=('img_path', 'img_id', 'ori_shape', 'img_shape', 'scale_factor', 'flip', 'flip_direction')
+)
 ]
 
 test_pipeline = [
@@ -55,7 +59,11 @@
         keep_ratio=True,
         bbox_format='xywh'  # 🔥 xywh format 명시
     ),
-    dict(type='PackDELIVERDetInputs')
+    # dict(type='PackDELIVERDetInputs')
+    dict(
+    type='PackDELIVERDetInputs',
+    meta_keys=('img_path', 'img_id', 'ori_shape', 'img_shape', 'scale_factor', 'flip', 'flip_direction', 'depth_path', 'lidar_path', 'thermal_path')
+)
 ]
 
 # Dataset configs
@@ -133,7 +141,7 @@
 ]
 
 # Training settings
-train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=24, val_interval=1)  # Updated for 2x
+train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=100, val_interval=1)  # Updated for 2x
 val_cfg = dict(type='ValLoop')
 test_cfg = dict(type='TestLoop')
 
@@ -144,7 +152,14 @@
     param_scheduler=dict(type='ParamSchedulerHook'),
     checkpoint=dict(type='CheckpointHook', interval=1, save_best='coco/bbox_mAP'),
     sampler_seed=dict(type='DistSamplerSeedHook'),
-    visualization=dict(type='DetVisualizationHook')
+    visualization=dict(
+        type='MultiModalVisualizationHook',
+        draw=True,
+        interval=1,
+        score_thr=0.3,
+        show=False,
+        test_out_dir='vis_results'
+    )
 )
 
 # Visualization settings

mmdet/engine/hooks/visualization_hook.py

@@ -556,12 +556,18 @@ def after_test_iter(self, runner: Runner, batch_idx: int, data_batch: dict,
         for data_sample in outputs:
             self._test_index += 1
             data_sample = data_sample.cpu()
-
+            print("data_sample:", data_sample)
             rgb_path = str(data_sample.img_path).strip()
-            depth_path = str(data_sample.metainfo.get('depth_path', '')).strip()
-            lidar_path = str(data_sample.metainfo.get('lidar_path', '')).strip()
-            thermal_path = str(data_sample.metainfo.get('thermal_path', '')).strip()
+            depth_path = str(data_sample.metainfo.get('depth_path') or 
+                            data_sample.metainfo.get('modality_paths', {}).get('depth', '')).strip()
+
+            lidar_path = str(data_sample.metainfo.get('lidar_path') or 
+                            data_sample.metainfo.get('modality_paths', {}).get('lidar', '')).strip()
+
+            thermal_path = str(data_sample.metainfo.get('thermal_path') or 
+                            data_sample.metainfo.get('modality_paths', {}).get('event', '')).strip()
 
+            print(rgb_path, depth_path, lidar_path, thermal_path)
             img_rgb = self._load_image(rgb_path)
             h, w, _ = img_rgb.shape
 
@@ -627,7 +633,10 @@ def draw_split(img, modality_shape):
                 pred_scaled = scale_bboxes(pred_bboxes, modality_shape, (h, w))
                 if pred_scaled is not None and pred_labels is not None:
                     self._visualizer.draw_bboxes(pred_scaled, edge_colors='red', alpha=0.8)
-                    label_texts = [classes[l] for l in pred_labels]
+                    label_texts = [
+                        f"{classes[l]} {round(s.item(), 2)}"
+                        for l, s in zip(pred_labels, pred_scores[keep])
+                    ]
                     self._visualizer.draw_texts(
                         label_texts,
                         pred_scaled[:, :2].int().numpy(),
@@ -639,6 +648,7 @@ def draw_split(img, modality_shape):
                             'pad': 0.7,
                             'edgecolor': 'none'
                         }] * len(pred_scaled))
+
                 img_pred = self._visualizer.get_image()
 
                 return np.concatenate([img_gt, img_pred], axis=1)
@@ -653,7 +663,13 @@ def draw_split(img, modality_shape):
 
             out_file = None
             if self.test_out_dir is not None:
-                out_file = osp.join(self.test_out_dir, f"{self._test_index:06d}.jpg")
+                basename = osp.basename(rgb_path)
+                basename = (basename
+                            .replace('_rgb_', '_')
+                            .replace('_depth_', '_')
+                            .replace('_event_', '_')
+                            .replace('_lidar_', '_'))
+                out_file = osp.join(self.test_out_dir, basename)
             if self.show:
                 self._visualizer.show(final_vis, win_name='multi_modal_split', wait_time=self.wait_time)
             if out_file is not None: