app.py

import os
import torch
import argparse
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt

from PIL import Image
from model import FoundModel
from misc import load_config
from torchvision import transforms as T

import gradio as gr

MAX_IM_SIZE = 512
NORMALIZE = T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
CACHE = True

def blend_images(bg, fg, alpha=0.5):
    bg = bg.convert('RGBA')
    fg = fg.convert('RGBA')
    blended = Image.blend(bg, fg, alpha=alpha)

    return blended


def predict(img_input):

    config = "configs/found_DUTS-TR.yaml"
    model_weights = "data/weights/decoder_weights.pt"

    # Configuration
    config = load_config(config)

    # ------------------------------------
    # Load the model
    model = FoundModel(vit_model=config.model["pre_training"],
                        vit_arch=config.model["arch"],
                        vit_patch_size=config.model["patch_size"],
                        enc_type_feats=config.found["feats"],
                        bkg_type_feats=config.found["feats"],
                        bkg_th=config.found["bkg_th"])
    # Load weights
    model.decoder_load_weights(model_weights)
    model.eval()
    print(f"Model {model_weights} loaded correctly.")

    # Load the image
    img_pil = Image.open(img_input)
    img = img_pil.convert("RGB")

    # Image transformations
    transforms = [T.ToTensor()]
    # Resize image if needed
    if img.size[0] > MAX_IM_SIZE or img.size[1] > MAX_IM_SIZE:
        transforms.append(T.Resize(MAX_IM_SIZE))
    transforms.append(NORMALIZE)
    t = T.Compose(transforms)

    img_t = t(img)[None,:,:,:]
    inputs = img_t
    
    # Forward step
    with torch.no_grad():
        preds, _, _, _ = model.forward_step(inputs, for_eval=True)

    # Apply FOUND
    sigmoid = nn.Sigmoid()
    h, w = img_t.shape[-2:]
    preds_up = F.interpolate(
        preds, scale_factor=model.vit_patch_size, mode="bilinear", align_corners=False
    )[..., :h, :w]
    preds_up = (
        (sigmoid(preds_up.detach()) > 0.5).squeeze(0).float()
    )

    return blend_images(img_pil.resize([img_t.shape[-1], img_t.shape[-2]]), T.ToPILImage()(preds_up))
    

title = 'FOUND - unsupervised object localization'
description = 'Gradio Demo for our CVPR23 paper "Unsupervised Object Localization: Observing the Background to Discover Objects"\n \
    The app is <i>running on CPUs</i>, inference times are therefore longer than those expected on GPU (80 FPS on a V100 GPU).\n \
    Please see below for more details.'
    
article = """
<h1 align="center">Unsupervised Object Localization: Observing the Background to Discover Objects</h1>
    
## Highlights
- Single **conv 1 x 1** layer trained to extract information from DINO [1] features.
- **No supervision**.
- Trained only for **2 epochs** on the dataset DUTS-TR.
- Inference runs at **80 FPS** on a V100 GPU.
- No post-processing applied in results here.

<i> Images provided are taken from VOC07 [2], ECSSD [3] and DUT-OMRON [4].</i>

## Citation
```
@inproceedings{simeoni2023found,
author    = {Siméoni, Oriane and Sekkat, Chloé and Puy, Gilles and Vobecky, Antonin and Zablocki, Éloi and Pérez, Patrick},
title     = {Unsupervised Object Localization: Observing the Background to Discover Objects},
booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition, {CVPR}},
year      = {2023},
}
```

### References

[1] M. Caron et al. Emerging properties in self-supervised vision transformers, ICCV 2021
[2] M. Everingham et al. The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results
[3] J. Shi et al. Hierarchical image saliency detection on extended CSSD, IEEE TPAMI 2016
[4] C. Yang et al. Saliency detection via graph-based manifold ranking, CVPR 2013

"""

examples = ["data/examples/VOC_000030.jpg",
            "data/examples/ECSSD_0010.png",
            "data/examples/VOC07_000038.jpg",
            "data/examples/VOC07_000075.jpg",
            "data/examples/DUT-OMRON_im103.png",
            ]


iface = gr.Interface(fn=predict,
                     title=title,
                     description=description,
                     article=article,
                     inputs=gr.Image(type='filepath'),
                     outputs=gr.Image(label="Unsupervised object localization", type="pil"),
                     examples=examples,
                     cache_examples=CACHE
                     )

iface.launch(show_error=True,
             enable_queue=True,
             inline=True,
             )