import gradio as gr """ ===================================================== Optical Flow: Predicting movement with the RAFT model ===================================================== Optical flow is the task of predicting movement between two images, usually two consecutive frames of a video. Optical flow models take two images as input, and predict a flow: the flow indicates the displacement of every single pixel in the first image, and maps it to its corresponding pixel in the second image. Flows are (2, H, W)-dimensional tensors, where the first axis corresponds to the predicted horizontal and vertical displacements. The following example illustrates how torchvision can be used to predict flows using our implementation of the RAFT model. We will also see how to convert the predicted flows to RGB images for visualization. """ import cv2 import numpy as np import os import sys import torch from PIL import Image import matplotlib.pyplot as plt import torchvision.transforms.functional as F from torchvision.io import read_video, read_image, ImageReadMode from torchvision.models.optical_flow import Raft_Large_Weights from torchvision.models.optical_flow import raft_large from torchvision.io import write_jpeg import torchvision.transforms as T import tempfile from pathlib import Path from urllib.request import urlretrieve from scipy.interpolate import LinearNDInterpolator from imageio import imread, imwrite from diffusers import StableDiffusionImg2ImgPipeline model_id = "runwayml/stable-diffusion-v1-5" pipe = StableDiffusionImg2ImgPipeline.from_pretrained(model_id, torch_dtype=torch.float16) pipe = pipe.to("cuda") def diffuse(source_img): source_img = Image.open(source_img) prompt = "space crew in a spaceship by Claude Monet" image = pipe(prompt, source_img).images[0] print(f"DIFFUSED IMG: {image}") image.save("astronaut_rides_horse.png") def write_flo(flow, filename): """ Write optical flow in Middlebury .flo format :param flow: optical flow map :param filename: optical flow file path to be saved :return: None from https://github.com/liruoteng/OpticalFlowToolkit/ """ # forcing conversion to float32 precision flow = flow.cpu().data.numpy() flow = flow.astype(np.float32) f = open(filename, 'wb') magic = np.array([202021.25], dtype=np.float32) (height, width) = flow.shape[0:2] w = np.array([width], dtype=np.int32) h = np.array([height], dtype=np.int32) magic.tofile(f) w.tofile(f) h.tofile(f) flow.tofile(f) f.close() def infer(): #video_url = "https://download.pytorch.org/tutorial/pexelscom_pavel_danilyuk_basketball_hd.mp4" #video_path = Path(tempfile.mkdtemp()) / "basketball.mp4" #_ = urlretrieve(video_url, video_path) #frames, _, _ = read_video(str("./spacex.mp4"), output_format="TCHW") #print(f"FRAME BEFORE stack: {frames[100]}") #testing difusion working properly diffuse("./frame1.jpg") input_frame_1 = read_image(str("./frame1.jpg"), ImageReadMode.UNCHANGED) print(f"FRAME 1: {input_frame_1}") input_frame_2 = read_image(str("./frame2.jpg"), ImageReadMode.UNCHANGED) print(f"FRAME 1: {input_frame_2}") #img1_batch = torch.stack([frames[0]]) #img2_batch = torch.stack([frames[1]]) img1_batch = torch.stack([input_frame_1]) img2_batch = torch.stack([input_frame_2]) print(f"FRAME AFTER stack: {img1_batch}") weights = Raft_Large_Weights.DEFAULT transforms = weights.transforms() def preprocess(img1_batch, img2_batch): img1_batch = F.resize(img1_batch, size=[520, 960]) img2_batch = F.resize(img2_batch, size=[520, 960]) return transforms(img1_batch, img2_batch) img1_batch, img2_batch = preprocess(img1_batch, img2_batch) print(f"shape = {img1_batch.shape}, dtype = {img1_batch.dtype}") #################################### # Estimating Optical flow using RAFT # ---------------------------------- # We will use our RAFT implementation from # :func:`~torchvision.models.optical_flow.raft_large`, which follows the same # architecture as the one described in the `original paper `_. # We also provide the :func:`~torchvision.models.optical_flow.raft_small` model # builder, which is smaller and faster to run, sacrificing a bit of accuracy. # If you can, run this example on a GPU, it will be a lot faster. device = "cuda" if torch.cuda.is_available() else "cpu" model = raft_large(weights=Raft_Large_Weights.DEFAULT, progress=False).to(device) model = model.eval() list_of_flows = model(img1_batch.to(device), img2_batch.to(device)) print(f"list_of_flows type = {type(list_of_flows)}") print(f"list_of_flows length = {len(list_of_flows)} = number of iterations of the model") #################################### # The RAFT model outputs lists of predicted flows where each entry is a # (N, 2, H, W) batch of predicted flows that corresponds to a given "iteration" # in the model. For more details on the iterative nature of the model, please # refer to the `original paper `_. Here, we # are only interested in the final predicted flows (they are the most acccurate # ones), so we will just retrieve the last item in the list. # # As described above, a flow is a tensor with dimensions (2, H, W) (or (N, 2, H, # W) for batches of flows) where each entry corresponds to the horizontal and # vertical displacement of each pixel from the first image to the second image. # Note that the predicted flows are in "pixel" unit, they are not normalized # w.r.t. the dimensions of the images. predicted_flows = list_of_flows[-1] print(f"predicted_flows dtype = {predicted_flows.dtype}") print(f"predicted_flows shape = {predicted_flows.shape} = (N, 2, H, W)") print(f"predicted_flows min = {predicted_flows.min()}, predicted_flows max = {predicted_flows.max()}") #################################### # Visualizing predicted flows # --------------------------- # Torchvision provides the :func:`~torchvision.utils.flow_to_image` utlity to # convert a flow into an RGB image. It also supports batches of flows. # each "direction" in the flow will be mapped to a given RGB color. In the # images below, pixels with similar colors are assumed by the model to be moving # in similar directions. The model is properly able to predict the movement of # the ball and the player. Note in particular the different predicted direction # of the ball in the first image (going to the left) and in the second image # (going up). from torchvision.utils import flow_to_image #flow_imgs = flow_to_image(predicted_flows) #print(flow_imgs) predicted_flow = list_of_flows[-1][0] print(f"predicted flow dtype = {predicted_flow.dtype}") print(f"predicted flow shape = {predicted_flow.shape}") flow_img = flow_to_image(predicted_flow).to("cpu") write_jpeg(flow_img, f"predicted_flow.jpg") flo_file = write_flo(predicted_flow, "flofile.flo") # define a transform to convert a tensor to PIL image transform = T.ToPILImage() # convert the tensor to PIL image using above transform #img = transform(frames[1]) img = transform(input_frame_2) img = img.resize((960, 520)) # display the PIL image #img.show() frame2pil = np.array(img.convert('RGB')) print(f"frame1pil: {frame2pil}") print(f"frame1pil shape: {frame2pil.shape}") print(f"frame1pil dtype: {frame2pil.dtype}") img.save('raw_frame2.jpg') numpy_array_flow = predicted_flow.permute(1, 2, 0).detach().cpu().numpy() print(f"numpy_array_flow: {numpy_array_flow}") print(f"numpy_array_flow shape: {numpy_array_flow.shape}") print(f"numpy_array_flow dtype: {numpy_array_flow.dtype}") h, w = numpy_array_flow.shape[:2] numpy_array_flow = numpy_array_flow.copy() numpy_array_flow[:, :, 0] + np.arange(w) numpy_array_flow[:, :, 1] + np.arange(h)[:, np.newaxis] # print('flow stats', flow.max(), flow.min(), flow.mean()) # print(flow) numpy_array_flow*1. # print('flow stats mul', flow.max(), flow.min(), flow.mean()) # res = cv2.remap(img, flow, None, cv2.INTER_LINEAR) res = cv2.remap(frame2pil, numpy_array_flow, None, cv2.INTER_LANCZOS4) print(res) res = Image.fromarray(res) res.save('wraped.jpg') blend2 = Image.open('raw_frame2.jpg') blend2 = Image.blend(res,blend2,0.5) blend2.save("blended2.jpg") return "done", "predicted_flow.jpg", ["flofile.flo"], 'raw_frame2.jpg', 'wraped.jpg', "blended2.jpg" #################################### # Bonus: Creating GIFs of predicted flows # --------------------------------------- # In the example above we have only shown the predicted flows of 2 pairs of # frames. A fun way to apply the Optical Flow models is to run the model on an # entire video, and create a new video from all the predicted flows. Below is a # snippet that can get you started with this. We comment out the code, because # this example is being rendered on a machine without a GPU, and it would take # too long to run it. # from torchvision.io import write_jpeg # for i, (img1, img2) in enumerate(zip(frames, frames[1:])): # # Note: it would be faster to predict batches of flows instead of individual flows # img1, img2 = preprocess(img1, img2) # list_of_flows = model(img1.to(device), img2.to(device)) # predicted_flow = list_of_flows[-1][0] # flow_img = flow_to_image(predicted_flow).to("cpu") # output_folder = "/tmp/" # Update this to the folder of your choice # write_jpeg(flow_img, output_folder + f"predicted_flow_{i}.jpg") #################################### # Once the .jpg flow images are saved, you can convert them into a video or a # GIF using ffmpeg with e.g.: # # ffmpeg -f image2 -framerate 30 -i predicted_flow_%d.jpg -loop -1 flow.gif gr.Interface(fn=infer, inputs=[], outputs=[gr.Textbox(), gr.Image(), gr.Files(), gr.Image(), gr.Image(), gr.Image()]).launch()