fix resnet divisible issue (#1658)

gluoncv/torch/model_zoo/pose/directpose_resnet_fpn.py

@@ -540,8 +540,8 @@ def forward(self, x):
         prev_features = self.lateral_convs[0](x[0])
         results.append(self.output_convs[0](prev_features))
         for features, lateral_conv, output_conv in zip(x[1:], self.lateral_convs[1:], self.output_convs[1:]):
-            top_down_features = F.interpolate(prev_features, scale_factor=2, mode="nearest")
             lateral_features = lateral_conv(features)
+            top_down_features = F.interpolate(prev_features, size=lateral_features.shape[-2:], mode="nearest")
             prev_features = lateral_features + top_down_features
             if self._fuse_type == "avg":
                 prev_features /= 2

scripts/pose/directpose/demo_directpose.py

@@ -18,6 +18,7 @@ def parse_args():
     parser.add_argument('--use-gpu', type=int, default=1)
     parser.add_argument("--image-width", type=int, default=1280)
     parser.add_argument("--image-height", type=int, default=736)
+    parser.add_argument("--score-threshold", type=float, default=0.5)
     parser.add_argument("--visualize", type=int, default=1, help="enable visualizer")
     parser.add_argument("--save-output", type=str, default='visualize_output.png', help='Save visualize result to image')
     parser.add_argument("--verbose", type=int, default=1)
@@ -77,6 +78,10 @@ def get_transforms():
         img_width=args.image_width, img_height=args.image_height)
     with torch.no_grad():
         predictions = net(images.to(device))[0]
+    if args.score_threshold > 0:
+        print(f"Applying score threshold: {args.score_threshold} to detected instances...")
+        valid_indices = predictions.scores > args.score_threshold
+        predictions = predictions[valid_indices]
     print(f"Detected {list(predictions.scores.size())} instances with scores: {predictions.scores}")
     # visualize
     if args.visualize or args.save_output: