Update to 2026-09-17 pipeline snapshot; add weights, L20 assets and recording via Git LFS

Source: RGB-D -> Dyn-HaMR -> L20 retargeting -> FoundationPose -> reference repair -> SPIDER, documented in docs/PIPELINE_LATEST.md and docs/SETUP_AND_WEIGHTS.md. Adds FoundationPose and nvdiffrast upstream snapshots, requirements/pipeline_venv.txt and the FoundationPose weight manifest/downloader. Assets (Git LFS): weights/ (WiLoR detector, HandFlow denoiser, UniDepth-L), FoundationPose checkpoints, HaMeR checkpoint, Dyn-HaMR HMP model and BMC constraints, L20 URDF/meshes, the 20260915_171525 D405 recording and the two box CADs. MANO models are not redistributed (third_party/hamer/_DATA/data/mano/README.txt). Environments, caches and run outputs excluded. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-17 11:27:46 +08:00
parent 7e4ef6f98b
commit ae28d55f81
761 changed files with 55598 additions and 59 deletions
@@ -0,0 +1,93 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+
+import os,sys
+import numpy as np
+code_dir = os.path.dirname(os.path.realpath(__file__))
+sys.path.append(code_dir)
+sys.path.append(f'{code_dir}/../../../../')
+from Utils import *
+import torch.nn.functional as F
+import torch
+import torch.nn as nn
+import cv2
+from functools import partial
+from network_modules import *
+from Utils import *
+
+
+
+class RefineNet(nn.Module):
+  def __init__(self, cfg=None, c_in=4, n_view=1):
+    super().__init__()
+    self.cfg = cfg
+    if self.cfg.use_BN:
+      norm_layer = nn.BatchNorm2d
+      norm_layer1d = nn.BatchNorm1d
+    else:
+      norm_layer = None
+      norm_layer1d = None
+
+    self.encodeA = nn.Sequential(
+      ConvBNReLU(C_in=c_in,C_out=64,kernel_size=7,stride=2, norm_layer=norm_layer),
+      ConvBNReLU(C_in=64,C_out=128,kernel_size=3,stride=2, norm_layer=norm_layer),
+      ResnetBasicBlock(128,128,bias=True, norm_layer=norm_layer),
+      ResnetBasicBlock(128,128,bias=True, norm_layer=norm_layer),
+    )
+
+    self.encodeAB = nn.Sequential(
+      ResnetBasicBlock(256,256,bias=True, norm_layer=norm_layer),
+      ResnetBasicBlock(256,256,bias=True, norm_layer=norm_layer),
+      ConvBNReLU(256,512,kernel_size=3,stride=2, norm_layer=norm_layer),
+      ResnetBasicBlock(512,512,bias=True, norm_layer=norm_layer),
+      ResnetBasicBlock(512,512,bias=True, norm_layer=norm_layer),
+    )
+
+    embed_dim = 512
+    num_heads = 4
+    self.pos_embed = PositionalEmbedding(d_model=embed_dim, max_len=400)
+
+    self.trans_head = nn.Sequential(
+      nn.TransformerEncoderLayer(d_model=embed_dim, nhead=num_heads, dim_feedforward=512, batch_first=True),
+		  nn.Linear(512, 3),
+    )
+
+    if self.cfg['rot_rep']=='axis_angle':
+      rot_out_dim = 3
+    elif self.cfg['rot_rep']=='6d':
+      rot_out_dim = 6
+    else:
+      raise RuntimeError
+    self.rot_head = nn.Sequential(
+      nn.TransformerEncoderLayer(d_model=embed_dim, nhead=num_heads, dim_feedforward=512, batch_first=True),
+		  nn.Linear(512, rot_out_dim),
+    )
+
+
+  def forward(self, A, B):
+    """
+    @A: (B,C,H,W)
+    """
+    bs = len(A)
+    output = {}
+
+    x = torch.cat([A,B], dim=0)
+    x = self.encodeA(x)
+    a = x[:bs]
+    b = x[bs:]
+
+    ab = torch.cat((a,b),1).contiguous()
+    ab = self.encodeAB(ab)  #(B,C,H,W)
+
+    ab = self.pos_embed(ab.reshape(bs, ab.shape[1], -1).permute(0,2,1))
+
+    output['trans'] = self.trans_head(ab).mean(dim=1)
+    output['rot'] = self.rot_head(ab).mean(dim=1)
+
+    return output