update

app.py

@@ -55,7 +55,7 @@ from lib.eval_utils.custom_utils import load_slam_cam
 from lib.vis.run_vis2 import lookat_matrix, run_vis2_on_video, run_vis2_on_video_cam
 from lib.vis.renderer_world import Renderer
 
-@spaces.GPU(duration=200)
+# @spaces.GPU(duration=200)
 def render_reconstruction(input_video, img_focal): 
     args = EasyDict()
     args.video_path = input_video
@@ -67,13 +67,23 @@ def render_reconstruction(input_video, img_focal):
     
     start_idx, end_idx, seq_folder, imgfiles = detect_track_video(args)
 
-    frame_chunks_all, img_focal = hawor_motion_estimation(args, start_idx, end_idx, seq_folder)
+    if os.path.exists(f'{seq_folder}/tracks_{start_idx}_{end_idx}/frame_chunks_all.npy'):
+        print("skip hawor motion estimation")
+        frame_chunks_all = joblib.load(f'{seq_folder}/tracks_{start_idx}_{end_idx}/frame_chunks_all.npy')
+        img_focal = args.img_focal
+    else:
+        frame_chunks_all, img_focal = hawor_motion_estimation(args, start_idx, end_idx, seq_folder)
 
     slam_path = os.path.join(seq_folder, f"SLAM/hawor_slam_w_scale_{start_idx}_{end_idx}.npz")
     if not os.path.exists(slam_path):
         hawor_slam(args, start_idx, end_idx)
     R_w2c_sla_all, t_w2c_sla_all, R_c2w_sla_all, t_c2w_sla_all = load_slam_cam(slam_path)
 
+    out_path = infiller_and_vis(args, start_idx, end_idx, frame_chunks_all, R_w2c_sla_all, t_w2c_sla_all, R_c2w_sla_all, t_c2w_sla_all, seq_folder, imgfiles)
+    return out_path
+
+@spaces.GPU(duration=80)
+def infiller_and_vis(args, start_idx, end_idx, frame_chunks_all, R_w2c_sla_all, t_w2c_sla_all, R_c2w_sla_all, t_c2w_sla_all, seq_folder, imgfiles):
     pred_trans, pred_rot, pred_hand_pose, pred_betas, pred_valid = hawor_infiller(args, start_idx, end_idx, frame_chunks_all)
 
     # vis sequence for this video

scripts/scripts_test_video/hawor_slam.py

@@ -11,6 +11,7 @@ from tqdm import tqdm
 import numpy as np
 import torch
 import cv2
+import spaces
 from PIL import Image
 from glob import glob
 from pycocotools import mask as masktool
@@ -43,6 +44,7 @@ def split_list_by_interval(lst, interval=1000):
     
     return start_indices, end_indices, split_lists
 
+@spaces.GPU(duration=80)
 def hawor_slam(args, start_idx, end_idx):
     # File and folders
     file = args.video_path

scripts/scripts_test_video/hawor_video.py

@@ -9,6 +9,7 @@ import cv2
 from tqdm import tqdm
 from glob import glob
 from natsort import natsorted
+import spaces
 
 from lib.pipeline.tools import parse_chunks, parse_chunks_hand_frame
 from lib.models.hawor import HAWOR
@@ -37,7 +38,7 @@ def load_hawor(checkpoint_path):
     return model, model_cfg
 
 
-
+@spaces.GPU(duration=80)
 def hawor_motion_estimation(args, start_idx, end_idx, seq_folder):
     model, model_cfg = load_hawor(args.checkpoint)
     device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')