1

app.py

@@ -80,7 +80,7 @@ def get_3D_model_from_scene(outdir, silent, scene, min_conf_thr=3, as_pointcloud
     return convert_scene_output_to_glb(outdir, rgbimg, pts3d, msk, focals, cams2world, as_pointcloud=as_pointcloud,
                                         transparent_cams=transparent_cams, cam_size=cam_size, show_cam=show_cam, silent=silent, save_name=save_name,
                                         cam_color=cam_color)
-                                        
+
 @spaces.GPU(duration=180)
 def generate_monocular_depth_maps(img_list, depth_prior_name):
     depth_list = []
@@ -120,9 +120,11 @@ def generate_monocular_depth_maps(img_list, depth_prior_name):
 def local_get_reconstructed_scene(filelist, min_conf_thr, as_pointcloud, mask_sky, clean_depth, transparent_cams, cam_size, depth_prior_name, **kw):
     depth_list, focallength_px_list = generate_monocular_depth_maps(filelist, depth_prior_name)
     imgs = load_images(filelist, depth_list, focallength_px_list, size=image_size, verbose=not silent,traj_format='custom', depth_prior_name=depth_prior_name)
-    pairs = []
-    pairs.append((imgs[0], imgs[1]))
-    pairs.append((imgs[1], imgs[0]))
+    # pairs = []
+    # pairs.append((imgs[0], imgs[1]))
+    # pairs.append((imgs[1], imgs[0]))
+    scenegraph_type = 'swinstride-5-noncyclic'
+    pairs = make_pairs(imgs, scene_graph=scenegraph_type, prefilter=None, symmetrize=True)
     if depth_prior_name == "Depth Pro":
       weights_path = "cyun9286/Align3R_DepthPro_ViTLarge_BaseDecoder_512_dpt"
     else:
@@ -130,8 +132,15 @@ def local_get_reconstructed_scene(filelist, min_conf_thr, as_pointcloud, mask_sk
     device = 'cuda' if torch.cuda.is_available() else 'cpu'
     model = AsymmetricCroCo3DStereo.from_pretrained(weights_path).to(device)  
     output = inference(pairs, model, device, batch_size=batch_size, verbose=not silent)
-    mode = GlobalAlignerMode.PairViewer
-    scene = global_aligner(output, device=device, mode=mode, verbose=not silent)
+    mode = GlobalAlignerMode.PointCloudOptimizer  
+    scene = global_aligner(output, device=device, mode=mode, verbose=not silent, shared_focal = True, temporal_smoothing_weight=0.01, translation_weight=1.0,
+                               flow_loss_weight=0.01, flow_loss_start_epoch=0.1, flow_loss_thre=25, use_self_mask=True,
+                               num_total_iter=300, empty_cache= len(filelist) > 72)
+    lr = 0.01
+    if mode == GlobalAlignerMode.PointCloudOptimizer:
+        loss = scene.compute_global_alignment(init='mst', niter=300, schedule='linear', lr=lr)
+    # mode = GlobalAlignerMode.PairViewer
+    # scene = global_aligner(output, device=device, mode=mode, verbose=not silent)
     save_folder = './output'
     os.makedirs(save_folder, exist_ok=True)
     outfile = get_3D_model_from_scene(save_folder, silent, scene, min_conf_thr, as_pointcloud, mask_sky, clean_depth, transparent_cams, cam_size)
@@ -141,7 +150,7 @@ def local_get_reconstructed_scene(filelist, min_conf_thr, as_pointcloud, mask_sk
 
 def run_example(snapshot, matching_conf_thr, min_conf_thr, cam_size, as_pointcloud, shared_intrinsics, filelist, **kw):
     return local_get_reconstructed_scene(filelist, cam_size, **kw)
-# filelist = ['/home/lipeng/ljh_code/Video_Depth_CVPR2025-main/Align3R/data/davis/DAVIS/JPEGImages/480p/bear/00000.jpg', '/home/lipeng/ljh_code/Video_Depth_CVPR2025-main/Align3R/data/davis/DAVIS/JPEGImages/480p/bear/00010.jpg']
+# filelist = ['/home/lipeng/ljh_code/Video_Depth_CVPR2025-main/Align3R/data/davis/DAVIS/JPEGImages/480p/bear/00000.jpg', '/home/lipeng/ljh_code/Video_Depth_CVPR2025-main/Align3R/data/davis/DAVIS/JPEGImages/480p/bear/00008.jpg','/home/lipeng/ljh_code/Video_Depth_CVPR2025-main/Align3R/data/davis/DAVIS/JPEGImages/480p/bear/00004.jpg', '/home/lipeng/ljh_code/Video_Depth_CVPR2025-main/Align3R/data/davis/DAVIS/JPEGImages/480p/bear/00010.jpg']
 # min_conf_thr = 1.1
 # as_pointcloud = True
 # mask_sky = False