python demo/image_demo.py images/airplane.jpg \
        configs/lae_dino/lae_dino_swin-t_pretrain_LAE-1M.py \
        --weights /path/to/model/ \
       --texts 'playground . road . tank . airplane . vehicle' -c \
       --palette random \
       --pred-score-thr 0.4

from PIL import Image
import re
import os
import json
import cv2
import copy
import random
import seaborn as sns

def plot_one_box(x, img, color=None, linewidth=1):
    # Plots one bounding box on image img
    tl = linewidth
    color = color or [random.randint(0, 255) for _ in range(3)]
    c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
    cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)


def get_seaborn_colors(n, palette='husl'):
    # 可选调色板: 'husl', 'Set2', 'dark', 'bright'等
    return sns.color_palette(palette, n_colors=n)

def inference(input: str, queries: list, thr: float, withlabel=True):  
    from mmdet.apis import DetInferencer
    init_args = {
        "model": "configs/lae_dino/lae_dino_swin-t_pretrain_LAE-1M.py",
        "weights": "checkpoint/lae_dino_swint_lae1m-28ca3a15.pth",
        "device": "cuda:0",
        "palette": "random",
    }
    inferencer = DetInferencer(**init_args)
    
    flag_folder = os.path.isdir(input)
    if flag_folder:
        output_name = os.path.basename(os.path.normpath(input))
    else:
        filename = os.path.basename(input)
        output_name, ext = os.path.splitext(filename)
        print(output_name, ext)
    
    inference_args = {
        "inputs": input,
        "texts": "",
        "custom_entities": True,
        "pred_score_thr": thr,
        "out_dir": "outputs/" + output_name,
        "no_save_pred" : False,
        "no_save_vis" : True,
    }

    if withlabel:
        # use inferencer
        inference_args["no_save_vis"] = False
    else:
        inference_args["no_save_vis"] = True

    texts = [re.sub('\n', '', line) for line in queries]
    if flag_folder:
        raise NotImplementedError # 暂时没写, MMDetection 可以直接推理整个文件夹, 不过似乎不能推理多提示词. 如果只有一个提示词, 那可以直接调用 inferencer
    else:
        origin_image = cv2.imread(inference_args["inputs"])
        inference_args["no_save_vis"] = not withlabel
        for text in texts:
            append_name = re.sub(r' \. ', '+', text)
            inference_args["texts"] = text
            inferencer(**inference_args)
            oldname = inference_args["out_dir"] + "/vis/" + output_name + ext
            newname = inference_args["out_dir"] + "/vis/" + output_name + append_name + ext
            oldnamep = inference_args["out_dir"] + "/preds/" + output_name + ".json"
            newnamep = inference_args["out_dir"] + "/preds/" + output_name + append_name + ".json"
            if withlabel:
                if (os.path.exists(newname)):
                    os.remove(newname)
                if (os.path.exists(newnamep)):
                   os.remove(newnamep)
                newname = inference_args["out_dir"] + "/vis/" + output_name + "_label_" + append_name + ext
                newnamep = inference_args["out_dir"] + "/preds/" + output_name + "_label_" + append_name + ".json"
                os.rename(oldname, newname)
                os.rename(oldnamep, newnamep)
            else:
                if (os.path.exists(newname)):
                    os.remove(newname)
                if (os.path.exists(newnamep)):
                    os.remove(newnamep)
                colors = get_seaborn_colors(len(text.split('.')))
                os.makedirs(inference_args["out_dir"] + "/vis", exist_ok=True)
                os.rename(oldnamep, newnamep)
                new_img = copy.copy(origin_image)
                with open(newnamep, 'r') as jsonfile:
                    args = json.load(jsonfile)
                    labels, scores, bboxes = args.values()
                    for iter in range(0, len(labels)):
                        if (scores[iter] >= thr):
                            bbox = bboxes[iter]
                            color = colors[labels[iter]]
                            plot_one_box(bbox, new_img, color, 1)
                    cv2.imwrite(newname, new_img)
                del new_img
        del origin_image

def main():
    config_name = "./config.txt"
    # line 1: picture x, y numbers, score_thr
    # line n: promptss
    with open(config_name, "r") as file:
        commands = file.readlines()
        _thr, yesno = commands[0].split(' ')
        thr = float(_thr)
        if yesno[0] == 'T':
            withlabel = True
        else:
            withlabel = False
        filename = re.sub('\n', '', commands[1])
        inference(filename, commands[2:], thr, withlabel)
if __name__ == "__main__":
    main()

0.05 True
<path to img>
car
car . vehicle . bus . a . al . ance . ish . ward