Merge pull request #25 from GradientSpaces/anchor-free

- Add anchor-free mode for RPF.
- Update Readme on camera-ready paper.

Author: suniique

README.md

@@ -17,6 +17,10 @@
 
 
 ## 🔔 News
+- [Oct 26, 2025] Our NeurIPS camera-ready [paper](https://arxiv.org/abs/2506.05282v2) is available! 🎉 
+  - We include additional experiments on generalizability and a new **anchor-free** model, which aligns more with practical assembly assumptions.
+  - We release **Version 1.1** to support the anchor-free model; see the [PR](https://github.com/GradientSpaces/Rectified-Point-Flow/pull/25) for more details.
+
 - [Sept 18, 2025] Our paper has been accepted to **NeurIPS 2025 (Spotlight)**; see you in San Diego!
 
 - [July 22, 2025] **Version 1.0**: We strongly recommend updating to this version, which includes:
@@ -292,6 +296,7 @@ Define parameters for Lightning's [Trainer](https://lightning.ai/docs/pytorch/la
 
 **Dataloader workers killed**: Usually this is a signal of insufficient CPU memory or stack. You may try to reduce the `num_workers`. 
 
+
 > [!NOTE]
 > Please don't hesitate to open an [issue](/issues) if you encounter any problems or bugs!
 

config/RPF_base_main_10k.yaml

@@ -23,3 +23,4 @@ ckpt_path: null          # when null, the checkpoint will be downloaded from Hug
 # Model settings
 model:
   compute_overlap_points: true
+  build_overlap_head: true

config/RPF_base_predict_overlap.yaml

@@ -23,6 +23,7 @@ hydra:
 
 model:
   compute_overlap_points: true
+  build_overlap_head: true
 
 data:
   limit_val_samples: 1000

config/RPF_base_pretrain.yaml

@@ -3,7 +3,8 @@ num_points_to_sample: 5000
 min_parts: 2
 max_parts: 64
 min_points_per_part: 20
-multi_anchor: true
+anchor_free: true
+multi_anchor: false
 
 data_root: ${data_root}
 dataset_names: ["ikea"]

config/data/ikea.yaml

@@ -3,7 +3,8 @@ num_points_to_sample: 5000
 min_parts: 2
 max_parts: 64
 min_points_per_part: 20
-multi_anchor: true
+anchor_free: true
+multi_anchor: false
 
 data_root: ${data_root}
 dataset_names: ["ikea", "partnet", "everyday", "twobytwo", "modelnet", "tudl"]

config/data/ikea_partnet_everyday_twobytwo_modelnet_tudl.yaml

@@ -3,7 +3,8 @@ num_points_to_sample: 5000
 min_parts: 2
 max_parts: 64
 min_points_per_part: 20
-multi_anchor: true
+anchor_free: true
+multi_anchor: false
 
 data_root: ${data_root}
 dataset_names: ["ikea", "partnet", "everyday", "twobytwo", "modelnet", "tudl", "objaverse_v1"]

config/data/ikea_partnet_everyday_twobytwo_modelnet_tudl_objverse.yaml

@@ -3,6 +3,7 @@ num_points_to_sample: 5000
 min_parts: 2
 max_parts: 64
 min_points_per_part: 20
+anchor_free: true
 multi_anchor: false
 
 data_root: ${data_root}

config/data/ikea_twobytwo.yaml

@@ -22,3 +22,4 @@ timestep_sampling: "u_shaped"
 inference_sampler: "euler"
 inference_sampling_steps: 50
 n_generations: 1
+anchor_free: true

config/model/rectified_point_flow.yaml

@@ -33,6 +33,7 @@ def __init__(
         up_axis: dict[str, str] = {},
         min_parts: int = 2,
         max_parts: int = 64,
+        anchor_free: bool = True,
         num_points_to_sample: int = 5000,
         min_points_per_part: int = 20,
         min_dataset_size: int = 2000,
@@ -51,6 +52,9 @@ def __init__(
                      If not provided, the up axis is assumed to be 'y'. This only affects the visualization.
             min_parts: Minimum number of parts in a point cloud.
             max_parts: Maximum number of parts in a point cloud.
+            anchor_free: Whether to use anchor-free mode.
+                     If True, the anchor part is centered and randomly rotated, like the non-anchor parts (default).
+                     If False, the anchor part is not centered and thus its pose in the CoM frame of the GT point cloud is given (align with GARF).
             num_points_to_sample: Number of points to sample from each point cloud.
             min_points_per_part: Minimum number of points per part.
             min_dataset_size: Minimum number of point clouds in a dataset.
@@ -65,6 +69,7 @@ def __init__(
         self.up_axis = up_axis
         self.min_parts = min_parts
         self.max_parts = max_parts
+        self.anchor_free = anchor_free
         self.num_points_to_sample = num_points_to_sample
         self.min_points_per_part = min_points_per_part
         self.batch_size = batch_size
@@ -120,6 +125,7 @@ def setup(self, stage: str):
                         num_points_to_sample=self.num_points_to_sample,
                         min_points_per_part=self.min_points_per_part,
                         min_dataset_size=self.min_dataset_size,
+                        anchor_free=self.anchor_free,
                         random_scale_range=self.random_scale_range,
                         multi_anchor=self.multi_anchor,
                     )
@@ -136,6 +142,7 @@ def setup(self, stage: str):
                         dataset_name=dataset_name,
                         min_parts=self.min_parts,
                         max_parts=self.max_parts,
+                        anchor_free=self.anchor_free,
                         num_points_to_sample=self.num_points_to_sample,
                         min_points_per_part=self.min_points_per_part,
                         limit_val_samples=self.limit_val_samples,
@@ -146,6 +153,7 @@ def setup(self, stage: str):
             logger.info(make_line())
             logger.info("Total Train Samples: " + str(self.train_dataset.cumulative_sizes[-1]))
             logger.info("Total Val Samples: " + str(self.val_dataset.cumulative_sizes[-1]))
+            logger.info("Anchor-free Mode: " + str(self.anchor_free))
 
         elif stage == "validate":
             self.val_dataset = ConcatDataset(
@@ -157,6 +165,7 @@ def setup(self, stage: str):
                         up_axis=self.up_axis.get(dataset_name, "y"),
                         min_parts=self.min_parts,
                         max_parts=self.max_parts,
+                        anchor_free=self.anchor_free,
                         num_points_to_sample=self.num_points_to_sample,
                         min_points_per_part=self.min_points_per_part,
                         limit_val_samples=self.limit_val_samples,
@@ -166,6 +175,7 @@ def setup(self, stage: str):
             )
             logger.info(make_line())
             logger.info("Total Val Samples: " + str(self.val_dataset.cumulative_sizes[-1]))
+            logger.info("Anchor-free Mode: " + str(self.anchor_free))
 
         elif stage in ["test", "predict"]:
             self.test_dataset = [
@@ -176,6 +186,7 @@ def setup(self, stage: str):
                     up_axis=self.up_axis.get(dataset_name, "y"),
                     min_parts=self.min_parts,
                     max_parts=self.max_parts,
+                    anchor_free=self.anchor_free,
                     num_points_to_sample=self.num_points_to_sample,
                     min_points_per_part=self.min_points_per_part,
                     limit_val_samples=self.limit_val_samples,
@@ -184,6 +195,7 @@ def setup(self, stage: str):
             ]
             logger.info(make_line())
             logger.info("Total Test Samples: " + str(sum(len(dataset) for dataset in self.test_dataset)))
+            logger.info("Anchor-free Mode: " + str(self.anchor_free))
 
     def train_dataloader(self):
         """Get training dataloader."""