Add files using upload-large-folder tool

Depth-Anything-3/da3_streaming/configs/tum.yaml

@@ -0,0 +1,51 @@
+Weights:
+  DA3: './weights/model.safetensors'
+  DA3_CONFIG: './weights/config.json'
+  SALAD: './weights/dino_salad.ckpt'
+
+Model:
+  chunk_size: 120
+  overlap: 60
+  loop_chunk_size: 20 # imgs of loop chunk = 2 * loop_chunk_size
+  loop_enable: True
+  useDBoW: False
+  delete_temp_files: True
+  align_lib: 'triton' # choose among 'triton' (GPU), 'torch' (GPU), 'numba' (CPU) or 'numpy' (CPU)
+  align_method: 'sim3' # choose among 'sim3', 'se3' or 'scale+se3'
+  scale_compute_method: 'auto' # choose among 'auto', 'ransac' or 'weighted'. It is used when align_method == 'scale+se3'
+  align_type: 'dense' # choose between 'dense' or 'sparse'. Sparse align is constructing, not avaible for now
+
+  ref_view_strategy: 'saddle_balanced' # 'middle', 'saddle_balanced', 'first' or 'saddle_sim_range'
+  ref_view_strategy_loop: 'saddle_balanced' # 'middle', 'saddle_balanced', 'first' or 'saddle_sim_range'
+  depth_threshold: 15.0
+
+  save_depth_conf_result: True
+  save_debug_info: False
+
+  Sparse_Align:
+    keypoint_select: 'orb' # choose among 'orb', 'fast' or 'disk'
+    keypoint_num: 5000
+
+  IRLS:
+    delta: 0.1
+    max_iters: 5
+    tol: 1e-9
+
+  Pointcloud_Save:
+    sample_ratio: 0.015
+    conf_threshold_coef: 0.75 # conf_threshold = np.mean(confs) * conf_threshold_coef
+
+Loop:
+  SALAD:
+    image_size: [336, 336]
+    batch_size: 32
+    similarity_threshold: 0.85
+    top_k: 5
+    use_nms: True
+    nms_threshold: 25
+
+  SIM3_Optimizer:
+    lang_version: 'cpp'
+    # choose between 'cpp' or 'python'. will auto set 'python' if c++ version has not installed
+    max_iterations: 30
+    lambda_init: 1e-6

Depth-Anything-3/da3_streaming/fastloop/__init__.py

@@ -0,0 +1,15 @@
+# Copyright (c) 2025 ByteDance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# Adapted from [VGGT-Long](https://github.com/DengKaiCQ/VGGT-Long)

Depth-Anything-3/da3_streaming/fastloop/solve.cpp

@@ -0,0 +1,111 @@
+// Copyright (c) 2025 ByteDance Ltd. and/or its affiliates
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// Adapted from [VGGT-Long](https://github.com/DengKaiCQ/VGGT-Long)
+
+#include <torch/extension.h>
+#include <vector>
+#include <unordered_map>
+#include <algorithm>
+#include <iostream>
+#include <Eigen/Core>
+#include <Eigen/Sparse>
+
+typedef Eigen::SparseMatrix<double> SpMat;
+typedef Eigen::Triplet<double> T;
+
+Eigen::VectorXd solve(const SpMat &A, const Eigen::VectorXd &b, int freen){
+
+  if (freen < 0){
+    const Eigen::SimplicialCholesky<SpMat> chol(A);
+    return chol.solve(b); // n x 1
+  }
+
+  const SpMat A_sub = A.topLeftCorner(freen, freen);
+  const Eigen::VectorXd b_sub = b.topRows(freen);
+  const Eigen::VectorXd delta = solve(A_sub, b_sub, -7);
+
+  Eigen::VectorXd delta2(b.rows());
+  delta2.setZero();
+  delta2.topRows(freen) = delta;
+
+  return delta2;
+}
+
+std::vector<torch::Tensor> solve_system(torch::Tensor J_Ginv_i, torch::Tensor J_Ginv_j, torch::Tensor ii, torch::Tensor jj, torch::Tensor res, float ep, float lm, int freen)
+{
+
+  const torch::Device device = res.device();
+  J_Ginv_i = J_Ginv_i.to(torch::kCPU);
+  J_Ginv_j = J_Ginv_j.to(torch::kCPU);
+  ii = ii.to(torch::kCPU);
+  jj = jj.to(torch::kCPU);
+  res = res.clone().to(torch::kCPU);
+
+  const int r = res.size(0);
+  const int n = std::max(ii.max().item<long>(), jj.max().item<long>()) + 1;
+
+  res.resize_({r*7});
+  float *res_ptr = res.data_ptr<float>();
+  Eigen::Map<Eigen::VectorXf> v(res_ptr, r*7);
+
+  SpMat J(r*7, n*7);
+  std::vector<T> tripletList;
+  tripletList.reserve(r*7*7*2);
+
+  auto ii_acc = ii.accessor<long,1>();
+  auto jj_acc = jj.accessor<long,1>();
+  auto J_Ginv_i_acc = J_Ginv_i.accessor<float,3>();
+  auto J_Ginv_j_acc = J_Ginv_j.accessor<float,3>();
+
+  for (int x=0; x<r; x++){
+    const int i = ii_acc[x];
+    const int j = jj_acc[x];
+    for (int k=0; k<7; k++){
+      for (int l=0; l<7; l++){
+        if (i == j)
+          exit(1);
+        const float val_i = J_Ginv_i_acc[x][k][l];
+        tripletList.emplace_back(x*7 + k, i*7 + l, val_i);
+        const float val_j = J_Ginv_j_acc[x][k][l];
+        tripletList.emplace_back(x*7 + k, j*7 + l, val_j);
+      }
+    }
+  }
+
+  J.setFromTriplets(tripletList.begin(), tripletList.end());
+  const SpMat Jt = J.transpose();
+  Eigen::VectorXd b = -(Jt * v.cast<double>());
+  SpMat A = Jt * J;
+
+  A.diagonal() += (A.diagonal() * lm);
+  A.diagonal().array() += ep;
+  Eigen::VectorXf delta = solve(A, b, freen*7).cast<float>();
+
+  torch::Tensor delta_tensor = torch::from_blob(delta.data(), {n*7}).clone().to(device);
+  delta_tensor.resize_({n, 7});
+  return {delta_tensor};
+
+  Eigen::Matrix<float, -1, -1, Eigen::RowMajor> dense_J(J.cast<float>());
+  torch::Tensor dense_J_tensor = torch::from_blob(dense_J.data(), {r*7, n*7}).clone().to(device);
+  dense_J_tensor.resize_({r, 7, n, 7});
+
+  return {delta_tensor, dense_J_tensor};
+
+}
+
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("solve_system", &solve_system, "temporal neighboor indicies");
+}

Depth-Anything-3/da3_streaming/fastloop/solve_python.py

@@ -0,0 +1,100 @@
+# Copyright (c) 2025 ByteDance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# Adapted from [VGGT-Long](https://github.com/DengKaiCQ/VGGT-Long)
+
+import numpy as np
+import torch
+from scipy.sparse import coo_matrix, csc_matrix
+from scipy.sparse.linalg import spsolve
+
+
+def solve_sparse(A: csc_matrix, b: np.ndarray, freen: int) -> np.ndarray:
+    """Solve linear system A * delta = b, supports submatrix solving"""
+    if freen < 0:
+        return spsolve(A, b)
+    else:
+        A_sub = A[:freen, :freen].tocsc()
+        b_sub = b[:freen]
+        delta_sub = spsolve(A_sub, b_sub)
+        delta = np.zeros_like(b)
+        delta[:freen] = delta_sub
+        return delta
+
+
+def solve_system_py(
+    J_Ginv_i: torch.Tensor,
+    J_Ginv_j: torch.Tensor,
+    ii: torch.Tensor,
+    jj: torch.Tensor,
+    res: torch.Tensor,
+    ep: float,
+    lm: float,
+    freen: int,
+) -> torch.Tensor:
+    # Ensure all tensors are on CPU
+    device = res.device
+    J_Ginv_i = J_Ginv_i.cpu()
+    J_Ginv_j = J_Ginv_j.cpu()
+    ii = ii.cpu()
+    jj = jj.cpu()
+    res = res.clone().cpu()
+
+    r = res.size(0)  # Number of edges
+    n = max(ii.max().item(), jj.max().item()) + 1  # Number of nodes
+
+    res_vec = res.view(-1).numpy().astype(np.float64)
+
+    rows, cols, data = [], [], []
+    ii_np = ii.numpy()
+    jj_np = jj.numpy()
+    J_Ginv_i_np = J_Ginv_i.numpy()
+    J_Ginv_j_np = J_Ginv_j.numpy()
+
+    for x in range(r):
+        i = ii_np[x]
+        j = jj_np[x]
+        if i == j:
+            raise ValueError("Self-edges are not allowed")
+
+        for k in range(7):
+            for l in range(7):
+                row_idx = x * 7 + k
+                col_idx_i = i * 7 + l
+                val_i = J_Ginv_i_np[x, k, l]
+                rows.append(row_idx)
+                cols.append(col_idx_i)
+                data.append(val_i)
+
+                col_idx_j = j * 7 + l
+                val_j = J_Ginv_j_np[x, k, l]
+                rows.append(row_idx)
+                cols.append(col_idx_j)
+                data.append(val_j)
+
+    J = coo_matrix((data, (rows, cols)), shape=(r * 7, n * 7)).tocsc()
+
+    b_vec = -J.T @ res_vec
+
+    A_mat = J.T @ J
+
+    diag = A_mat.diagonal()
+    new_diag = diag * (1.0 + lm) + ep
+    A_mat.setdiag(new_diag)
+
+    freen_total = freen * 7
+    delta = solve_sparse(A_mat.tocsc(), b_vec, freen_total)
+
+    delta_tensor = torch.from_numpy(delta.astype(np.float32)).view(n, 7).to(device)
+    return delta_tensor

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/affine_transformation_of_depth/02.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59ef74b7185341e52ffb2612008ef0de5fe4c5f616ee4a460e4812074675e7de
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/affine_transformation_of_depth/03.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24e4686b9465730f0fca137ce2b6470a7a51e886a5a1358ab2bffef216d0f87e
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/boundary/01.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ad23521132eb0a78c32cc32327d72bcdd24116c43ef5c525ee6c89f6c15b078
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/boundary/02.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa3b37afe04ca9cee0d9def8257e34cd3c45dcf9a4196c1f660ed6e18f40975c
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/boundary/03.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6a17c94c0342f27d647b412796dc0fc8de271729cd0325856bb5366dc60e8fe
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/boundary/06.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d08d17dd8e08c54879510417689a6ac7b04eadba08a5deb4870401fa853d707
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/boundary/09.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1ad664a424374422e779e13346b626b585b689aa75eaf5ab25faf4b8fe4d3c4
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/boundary/12.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af4565beaacca75ca028fcc3f66de988f6e0f4e66d6034f37defcc3fedcdadb4
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/curv_high_freq/0.002_1.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a8b81dc4ee7d3c95e108b0830bebd140db24adca2838e8df7929bdd930ff12c
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/curv_high_freq/0.004_1.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f2ef03a831fef69cad4a8c6e275b0655f15baf0fda7c5f1194c52d7bf934d26
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/curv_high_freq/0.006_1.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e53055694ccdf2f4c06750711c69671f136c34246e3e0d8a4963e9680fdb900
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/curv_high_freq/0.008_1.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:711dece9ffd63d0f6f47256e25e7a3eb617c7899cf74df0e049ec818e0e23ae0
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/curv_high_freq/0.010_1.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc586307efab8ebf4e723f39d7b46e93a38b870be55c61b8956fdff0c40b86fe
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/curv_high_freq/0.015_1.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a1de897b39a662ded584aec1cb63df2b00e2ebb5fd6a68a2d79fd77c86b022c
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/indoor__slow_solve-long_focal-2025-09-06-19-53-32___142efe82/synth_depth/curv_high_freq/0.020_1.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22444e9dd24c0885f07f2d419646f5d0b1f9ad9b6522e5e48e62d89f34ba85ed
+size 4608778

data/infinigen/test_scenes_release_cleaned_final/nature__desert-long_focal-2025-09-07-11-43-28___1c33a692/synth_depth/curv_high_freq/0.006_1.0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9884661e49e1a3801e72cd448779fce89d806fbed349c2a2e295d5136bd6b00
+size 4608778