Skip to content

Latest commit

 

History

History
154 lines (107 loc) · 5.54 KB

README.md

File metadata and controls

154 lines (107 loc) · 5.54 KB
Raw

MIMO-hack

https://github.com/Uminosachi/inpaint-anything

test_motion.py

https://huggingface.co/lilpotat/pytorch3d/tree/main

Dataset

We create a human video dataset called HUD-7K to train our model. This dataset consists of 5K real character videos and 2K synthetic character animations. The former does not require any annotations and can be automatically decomposed to various spatial attributes via our scheme. To enlarge the range of the real dataset, we also synthesize 2K videos by rendering character animations in complex motions under multiple camera views, utilizing En3D [21]. These synthetic videos are equipped with accurate annotations due to completely controlled production.

https://github.com/menyifang/En3D

Synthetic Training data

https://openxlab.org.cn/datasets/OpenXDLab/SynBody

pip install openxlab #Install

pip install -U openxlab #Upgrade

openxlab login # Log in and enter the corresponding AK/SK. Please view AK/SK at usercenter

openxlab dataset info --dataset-repo OpenXDLab/SynBody # Dataset information viewing and View Dataset File List

openxlab dataset get --dataset-repo OpenXDLab/SynBody #Dataset download

openxlab dataset download --dataset-repo OpenXDLab/SynBody --source-path /README.md --target-path /path/to/local/folder #Dataset file download

Julia Models - neutral

download this https://smpl-x.is.tue.mpg.de/download.php

Sapiens - get image - produce depths / normals / pose

https://github.com/facebookresearch/sapiens

python pose_vis.py '/home/oem/Desktop/image_1.png'  test.png output.json
python normal_vis.py '/home/oem/Desktop/image_1.png'  test.png 
python depth_estimation.py input_image.png output_depth_image.png output_depth_map.npy --depth_model 1b --seg_model fg-bg-1b

LAMA - SOTA inpainting

https://github.com/advimman/lama

Todo Components

1. Setup and Dependencies

  • Import necessary libraries for deep learning and 3D processing
    • torch, torchvision, pytorch3d, clip, SMPL, diffusers
    • Custom utility functions for video loading, depth estimation, and mask computation

2. Define Model Components

2.1 Temporal Attention Layer

  • Implement temporal attention layer for improved 3D motion representation
    • Utilize GroupNorm and BasicTransformerBlock from diffusers

2.2 Differentiable Rasterizer

  • Define DifferentiableRasterizer for projecting 3D models into 2D feature maps
    • Use pytorch3d for rendering and rasterization
    • Integrate PerspectiveCameras for basic 3D-to-2D projection

2.3 Structured Motion Encoder

  • Create StructuredMotionEncoder for encoding SMPL-based human motion
    • Load SMPL model for 3D human body modeling
    • Project SMPL vertices onto 2D plane
    • Encode motion using 3D CNN layers

2.4 Canonical Identity Encoder

  • Develop CanonicalIdentityEncoder for disentangling identity attributes
    • Use CLIP model for global and local feature extraction
    • Add a custom reference network for additional local features

2.5 Scene and Occlusion Encoder

  • Implement SceneOcclusionEncoder using a shared pre-trained VAE for encoding
    • Add temporal convolution layers for time-series input processing

2.6 Diffusion Decoder

  • Design the DiffusionDecoder using UNet3DConditionModel from diffusers
    • Customize with temporal attention blocks and stable diffusion-based decoder

2.7 MIMO Model

  • Combine all components into a unified MIMOModel
    • Motion encoder, identity encoder, and scene/occlusion encoder as input conditions
    • Use diffusion decoder for video synthesis from latent representations

3. Dataset Handling

3.1 Dataset Class

  • Create MIMODataset class for loading video data and corresponding attributes
    • Load video frames, SMPL parameters, identity images, and camera parameters
    • Use LAMA inpainting for scene reconstruction

3.2 Data Preprocessing

  • Implement functions for data preprocessing
    • Load video frames as tensors
    • Estimate depth using pre-trained MiDaS or Sapiens depth estimator
    • Detect and track humans using Detectron2
    • Extract SMPL parameters for human pose estimation
    • Inpaint scene using OpenCV inpainting

3.3 Mask Computation

  • Compute masks for human, scene, and occlusion layers
    • Use depth maps and human masks for spatial decomposition
    • Remove small components in masks using connected component analysis

4. Training Procedure

4.1 Forward Diffusion Sampling

  • Define the forward diffusion process with noise scheduling
    • Implement the linear_beta_schedule for noise addition

4.2 Training Loop

  • Implement the main training loop for the MIMO model
    • Load data from MIMODataset
    • Apply noise to latent codes and predict the noise residual using the diffusion decoder
    • Optimize with MSE loss

5. Inference Pipeline

  • Implement the inference function for generating new character videos
    • Start from random noise and use DDIM scheduler for step-wise denoising
    • Encode conditions (identity, motion, scene) and generate a video from latent space

6. Hyperparameters and Configuration

  • Define hyperparameters for training and inference
    • Number of vertices, feature dimension, image size, timesteps, batch size, learning rate

7. Additional Components

  • Add utility functions for processing video frames, depth maps, and masks
    • load_video, estimate_depth, detect_and_track_humans, inpaint_scene, compute_masks

8. Saving and Loading the Model

  • Save trained model weights using torch.save
  • Load the saved model for inference