app.py

import os.path as osp
import gradio as gr
import torch
import logging
import torchvision
from torchvision.models.detection.faster_rcnn import fasterrcnn_resnet50_fpn
from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
from src.detection.graph_utils import add_bbox
from src.detection.vision import presets
logging.getLogger('PIL').setLevel(logging.CRITICAL)


def load_model(baseline: bool = False):
    if baseline:
        model = fasterrcnn_resnet50_fpn(
            weights="DEFAULT")
    else:
        model = fasterrcnn_resnet50_fpn()
        in_features = model.roi_heads.box_predictor.cls_score.in_features
        model.roi_heads.box_predictor = FastRCNNPredictor(in_features, 2)
        checkpoint = torch.load(
            "model_split_3_FT_MOT17.pth", map_location="cpu")
        model.load_state_dict(checkpoint["model"])
    device = torch.device('cuda:0')
    model.to(device)
    model.eval()
    return model


def frcnn_motsynth(image):
    model = load_model(baseline=True)
    transformEval = presets.DetectionPresetEval()
    image_tensor = transformEval(image, None)[0]
    prediction = model([image_tensor])[0]
    image_w_bbox = add_bbox(image_tensor, prediction, 0.80)
    torchvision.io.write_png(image_w_bbox, "custom_out.png")
    return "custom_out.png"


def frcnn_coco(image):
    model = load_model(baseline=True)
    transformEval = presets.DetectionPresetEval()
    image_tensor = transformEval(image, None)[0]
    prediction = model([image_tensor])[0]
    image_w_bbox = add_bbox(image_tensor, prediction, 0.80)
    torchvision.io.write_png(image_w_bbox, "baseline_out.png")
    return "baseline_out.png"


title = "Domain shift adaption on pedestrian detection with Faster R-CNN"
description = "![alt text](http://www.aiacademy.unimore.it/media/news/ai-logo-white_2ND_EDITION.png)"
examples = "/input_examples"

io_baseline = gr.Interface(frcnn_coco, gr.Image(type="pil"), gr.Image(
    type="file", shape=(1920, 1080), label="Baseline Model trained on COCO + FT on MOT17"))

io_custom = gr.Interface(frcnn_motsynth, gr.Image(type="pil"), gr.Image(
    type="file", shape=(1920, 1080), label="Faster R-CNN trained on MOTSynth + FT on MOT17"))

gr.Parallel(io_baseline, io_custom, title=title,
            description=description, examples=examples).launch(enable_queue=True)