app.py

import cv2
import torch
import numpy as np
from transformers import DPTForDepthEstimation, DPTImageProcessor
import gradio as gr

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-swinv2-tiny-256", torch_dtype=torch.float16).to(device)
processor = DPTImageProcessor.from_pretrained("Intel/dpt-swinv2-tiny-256")

color_map = cv2.applyColorMap(np.arange(256, dtype=np.uint8), cv2.COLORMAP_INFERNO)

input_tensor = torch.zeros((1, 3, 128, 128), dtype=torch.float16, device=device)
depth_map = np.zeros((128, 128), dtype=np.float32)
depth_map_colored = np.zeros((128, 128, 3), dtype=np.uint8)

def preprocess_image(image):
    return cv2.resize(image, (128, 128), interpolation=cv2.INTER_AREA).transpose(2, 0, 1).astype(np.float32) / 255.0

@torch.inference_mode()
def process_frame(image):
    preprocessed = preprocess_image(image)
    input_tensor[0] = torch.from_numpy(preprocessed).to(device)

    if torch.cuda.is_available():
        torch.cuda.synchronize()

    predicted_depth = model(input_tensor).predicted_depth
    np.subtract(predicted_depth.squeeze().cpu().numpy(), predicted_depth.min().item(), out=depth_map)
    np.divide(depth_map, depth_map.max(), out=depth_map)
    np.multiply(depth_map, 255, out=depth_map)
    depth_map = depth_map.astype(np.uint8)
    cv2.applyColorMap(depth_map, color_map, dst=depth_map_colored)
    
    return depth_map_colored

interface = gr.Interface(
    fn=process_frame,
    inputs=gr.Image(source="webcam", streaming=True),
    outputs="image",
    live=True,
    refresh_rate=0.1
)

interface.launch()