release dataset

LASO-C.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3de34d3940f38b58e723b717581c9d1a6a24ccc3e3a88bfb5b8960b126323c5e
+size 2528938849

PIAD-C.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a39c5adf3c24fdef21aec0043b407497acb45ab1f680e2d1f12ec43cba56ca5
+size 2605080518

README.md

@@ -0,0 +1,67 @@
+<p align="center">
+  
+  <h3 align="center"><strong>GEAL: Generalizable 3D Affordance Learning with Cross-Modal Consistency</strong></h3>
+
+  <p align="center">
+      <a href="https://dylanorange.github.io" target='_blank'>Dongyue Lu</a>&nbsp;&nbsp;&nbsp;
+      <a href="https://ldkong.com" target='_blank'>Lingdong Kong</a>&nbsp;&nbsp;&nbsp;
+      <a href="https://tianxinhuang.github.io/" target='_blank'>Tianxin Huang</a>&nbsp;&nbsp;&nbsp;
+      <a href="https://www.comp.nus.edu.sg/~leegh/">Gim Hee Lee</a>&nbsp;&nbsp;&nbsp;
+    </br>
+  National University of Singapore&nbsp;&nbsp;&nbsp;
+  </p>
+
+</p>
+
+<p align="center">
+  <a href="https://dylanorange.github.io/projects/geal/static/files/geal.pdf" target='_blank'>
+    <img src="https://img.shields.io/badge/Paper-%F0%9F%93%83-lightblue">
+  </a>
+  
+  <a href="https://dylanorange.github.io/projects/geal" target='_blank'>
+    <img src="https://img.shields.io/badge/Project-%F0%9F%94%97-blue">
+  </a>
+
+  <a href="https://huggingface.co/datasets/dylanorange/geal" target="_blank">
+    <img src="https://img.shields.io/badge/Dataset-%20Hugging%20Face-yellow">
+</a>
+
+
+</p>
+
+
+## About 🛠️
+
+**GEAL** is a novel framework designed to enhance the generalization and robustness of 3D affordance learning by leveraging pre-trained 2D models.
+
+To facilitate robust 3D affordance learning across diverse real-world scenarios, we establish two 3D affordance robustness benchmarks: **PIAD-C** and **LASO-C**, based on the test sets of the commonly used datasets PIAD and LASO. We apply seven types of corruptions:
+
+- **Add Global**
+- **Add Local**
+- **Drop Global**
+- **Drop Local**
+- **Rotate**
+- **Scale**
+- **Jitter**
+
+Each corruption is applied with five severity levels, resulting in a total of **4890 object-affordance pairings**, comprising **17 affordance categories** and **23 object categories** with **2047 distinct object shapes**.
+
+
+<div style="text-align: center;">
+    <img src="supp_benchmark_1.jpg" alt="GEAL Performance GIF" style="max-width: 100%; height: auto; width: 1000px;">
+    <img src="supp_benchmark_2.jpg" alt="GEAL Performance GIF" style="max-width: 100%; height: auto; width: 1000px;">
+</div>
+
+## Updates 📰
+
+- **[2024.12]** - We have released our **PIAD-C** and **LASO-C** datasets! 🎉📂
+
+
+## Dataset and Code Release 🚀
+
+We are excited to announce the release of our dataset and dataloader:
+
+- **Dataset**: Available in the `PIAD-C` and `LASO-C` files 📜
+- **Dataloader**: Available in the `dataset.py` file 📜
+
+Stay tuned! Further evaluation code will be coming soon. 🔧✨

dataset.py

@@ -0,0 +1,72 @@
+import os
+import pandas as pd
+import pickle
+import numpy as np
+from torch.utils.data import Dataset
+
+CLASSES = ["Bag", "Bed", "Bowl","Clock", "Dishwasher", "Display", "Door", "Earphone", "Faucet",
+    "Hat", "StorageFurniture", "Keyboard", "Knife", "Laptop", "Microwave", "Mug",
+    "Refrigerator", "Chair", "Scissors", "Table", "TrashCan", "Vase", "Bottle"]
+
+AFFORD_CL = ['lay','sit','support','grasp','lift','contain','open','wrap_grasp','pour', 
+                'move','display','push','pull','listen','wear','press','cut','stab']
+
+def pc_normalize(pc):
+    centroid = np.mean(pc, axis=0)
+    pc = pc - centroid
+    m = np.max(np.sqrt(np.sum(pc**2, axis=1)))
+    pc = pc / m
+    return pc, centroid, m
+
+class Corrupt(Dataset):
+
+    def __init__(self,
+                 corrupt_type='scale',
+                 level=0
+                 ):
+        
+        #replace with the path to the LASO-C/PIAD-C dataset
+        data_root='LASO-C'
+
+        file_name = f'{corrupt_type}_{level}.pkl'
+
+        self.corrupt_type = corrupt_type
+        self.level = level
+
+        self.cls2idx = {cls.lower():np.array(i).astype(np.int64) for i, cls in enumerate(CLASSES)}
+        self.aff2idx = {cls:np.array(i).astype(np.int64) for i, cls in enumerate(AFFORD_CL)}
+
+        with open(os.path.join(data_root, 'point', file_name), 'rb') as f:
+            self.anno = pickle.load(f)
+
+        self.question_df = pd.read_csv(os.path.join(data_root, 'text', 'Affordance-Question.csv'))
+       
+    def find_rephrase(self, df, object_name, affordance):
+
+        qid = 'Question0'
+        result = df.loc[(df['Object'] == object_name) & (df['Affordance'] == affordance), [qid]]
+        if not result.empty:
+            return result.iloc[0][qid]
+        else:
+            raise NotImplementedError
+            
+    def __getitem__(self, index):
+
+        data = self.anno[index]        
+        cls = data['class']
+        affordance = data['affordance']
+        gt_mask = data['mask']
+        point_set = data['point']
+        point_set,_,_ = pc_normalize(point_set)    
+
+        question = self.find_rephrase(self.question_df, cls, affordance)
+        
+        affordance = self.aff2idx[affordance]
+
+        point_input = point_set.transpose()
+
+        return point_input, self.cls2idx[cls], gt_mask, question, affordance
+
+    def __len__(self):
+
+        return len(self.anno)