【AI实战营第二期】第三次作业——基于 RTMDet 的气球检测（包含数据集）

作业：基于 RTMDet 的气球检测

背景：熟悉目标检测和 MMDetection 常用自定义流程。

任务：

基于提供的 notebook，将 cat 数据集换成气球数据集;
按照视频中 notebook 步骤，可视化数据集和标签;
使用MMDetection算法库，训练 RTMDet 气球目标检测算法，可以适当调参，提交测试集评估指标;
用网上下载的任意包括气球的图片进行预测，将预测结果发到群里;
按照视频中 notebook 步骤，对 demo 图片进行特征图可视化和 Box AM 可视化，将结果发到群里
需提交的测试集评估指标（不能低于baseline指标的50%）
目标检测 RTMDet-tiny 模型性能 不低于65 mAP。
数据集
气球数据集可以直接下载https://download.openmmlab.com/mmyolo/data/balloon_dataset.zip

P.S. 同时也欢迎各位选择更复杂的数据集进行训练，如选用同济子豪兄的十类饮料目标检测数据集Drink_28

展示图片

使用pyplot展示训练集的图片，代码如下：

# 数据集可视化

import os
import matplotlib.pyplot as plt
from PIL import Image

original_images = []
images = []
texts = []
plt.figure(figsize=(16, 5))

image_paths = [filename for filename in os.listdir('balloon/train')][:8]

for i, filename in enumerate(image_paths):
    name = os.path.splitext(filename)[0]

    image = Image.open('balloon/train/' + filename).convert("RGB")

    plt.subplot(2, 4, i + 1)
    plt.imshow(image)
    plt.title(f"{filename}")
    plt.xticks([])
    plt.yticks([])

plt.tight_layout()
plt.show()

由于数据集不是标椎的COCO数据集，所以先将格式转为COCO

import os.path as osp

import mmcv

from mmengine.fileio import dump, load
from mmengine.utils import track_iter_progress


def convert_balloon_to_coco(ann_file, out_file, image_prefix):
    data_infos = load(ann_file)

    annotations = []
    images = []
    obj_count = 0
    for idx, v in enumerate(track_iter_progress(data_infos.values())):
        filename = v['filename']
        img_path = osp.join(image_prefix, filename)
        height, width = mmcv.imread(img_path).shape[:2]

        images.append(
            dict(id=idx, file_name=filename, height=height, width=width))

        for _, obj in v['regions'].items():
            assert not obj['region_attributes']
            obj = obj['shape_attributes']
            px = obj['all_points_x']
            py = obj['all_points_y']
            poly = [(x + 0.5, y + 0.5) for x, y in zip(px, py)]
            poly = [p for x in poly for p in x]

            x_min, y_min, x_max, y_max = (min(px), min(py), max(px), max(py))

            data_anno = dict(
                image_id=idx,
                id=obj_count,
                category_id=0,
                bbox=[x_min, y_min, x_max - x_min, y_max - y_min],
                area=(x_max - x_min) * (y_max - y_min),
                segmentation=[poly],
                iscrowd=0)
            annotations.append(data_anno)
            obj_count += 1

    coco_format_json = dict(
        images=images,
        annotations=annotations,
        categories=[{
            'id': 0,
            'name': 'balloon'
        }])
    dump(coco_format_json, out_file)


if __name__ == '__main__':
    convert_balloon_to_coco(ann_file='balloon/train/via_region_data.json',
                            out_file='balloon/train/annotation_coco.json',
                            image_prefix='balloon/train')
    convert_balloon_to_coco(ann_file='balloon/val/via_region_data.json',
                            out_file='balloon/val/annotation_coco.json',
                            image_prefix='balloon/val')

然后，展示COCO格式的数据集，代码如下：

from pycocotools.coco import COCO
import numpy as np
import os.path as osp
from matplotlib.collections import PatchCollection
from matplotlib.patches import Polygon
import matplotlib.pyplot as plt
from PIL import Image

def apply_exif_orientation(image):
    _EXIF_ORIENT = 274
    if not hasattr(image, 'getexif'):
        return image

    try:
        exif = image.getexif()
    except Exception:
        exif = None

    if exif is None:
        return image

    orientation = exif.get(_EXIF_ORIENT)

    method = {
        2: Image.FLIP_LEFT_RIGHT,
        3: Image.ROTATE_180,
        4: Image.FLIP_TOP_BOTTOM,
        5: Image.TRANSPOSE,
        6: Image.ROTATE_270,
        7: Image.TRANSVERSE,
        8: Image.ROTATE_90,
    }.get(orientation)
    if method is not None:
        return image.transpose(method)
    return image


def show_bbox_only(coco, anns, show_label_bbox=True, is_filling=True):
    """Show bounding box of annotations Only."""
    if len(anns) == 0:
        return

    ax = plt.gca()
    ax.set_autoscale_on(False)

    image2color = dict()
    for cat in coco.getCatIds():
        image2color[cat] = (np.random.random((1, 3)) * 0.7 + 0.3).tolist()[0]

    polygons = []
    colors = []

    for ann in anns:
        color = image2color[ann['category_id']]
        bbox_x, bbox_y, bbox_w, bbox_h = ann['bbox']
        poly = [[bbox_x, bbox_y], [bbox_x, bbox_y + bbox_h],
                [bbox_x + bbox_w, bbox_y + bbox_h], [bbox_x + bbox_w, bbox_y]]
        polygons.append(Polygon(np.array(poly).reshape((4, 2))))
        colors.append(color)

        if show_label_bbox:
            label_bbox = dict(facecolor=color)
        else:
            label_bbox = None

        ax.text(
            bbox_x,
            bbox_y,
            '%s' % (coco.loadCats(ann['category_id'])[0]['name']),
            color='white',
            bbox=label_bbox)

    if is_filling:
        p = PatchCollection(
            polygons, facecolor=colors, linewidths=0, alpha=0.4)
        ax.add_collection(p)
    p = PatchCollection(
        polygons, facecolor='none', edgecolors=colors, linewidths=2)
    ax.add_collection(p)


coco = COCO('balloon/val/annotation_coco.json')
image_ids = coco.getImgIds()
print(image_ids)
np.random.shuffle(image_ids)

plt.figure(figsize=(16, 5))

# 只可视化 8 张图片
for i in range(8):
    image_data = coco.loadImgs(image_ids[i])[0]
    image_path = osp.join('balloon/val/', image_data['file_name'])
    annotation_ids = coco.getAnnIds(
        imgIds=image_data['id'], catIds=[], iscrowd=0)
    annotations = coco.loadAnns(annotation_ids)

    ax = plt.subplot(2, 4, i + 1)
    image = Image.open(image_path).convert("RGB")

    # 这行代码很关键，否则可能图片和标签对不上
    image = apply_exif_orientation(image)

    ax.imshow(image)

    show_bbox_only(coco, annotations)

    plt.title(f"{image_data['file_name']}")
    plt.xticks([])
    plt.yticks([])

plt.tight_layout()
plt.show()

配置文件

编写在根目录，新建rtmdet_tiny_1xb12-40e_cat.py，插入代码：

_base_ = 'configs/rtmdet/rtmdet_tiny_8xb32-300e_coco.py'

data_root = '.'

# 非常重要
metainfo = {
    # 类别名，注意 classes 需要是一个 tuple，因此即使是单类，
    # 你应该写成 `cat,` 很多初学者经常会在这犯错
    'classes': ('balloon',),
    'palette': [
        (220, 20, 60),
    ]
}
num_classes = 1

# 训练 40 epoch
max_epochs = 40
# 训练单卡 bs= 12
train_batch_size_per_gpu = 12
# 可以根据自己的电脑修改
train_num_workers = 4

# 验证集 batch size 为 1
val_batch_size_per_gpu = 1
val_num_workers = 2

# RTMDet 训练过程分成 2 个 stage，第二个 stage 会切换数据增强 pipeline
num_epochs_stage2 = 5

# batch 改变了，学习率也要跟着改变， 0.004 是 8卡x32 的学习率
base_lr = 12 * 0.004 / (32*8)

# 采用 COCO 预训练权重
load_from = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/rtmdet_tiny_8xb32-300e_coco/rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth'  # noqa

model = dict(
    # 考虑到数据集太小，且训练时间很短，我们把 backbone 完全固定
    # 用户自己的数据集可能需要解冻 backbone
    backbone=dict(frozen_stages=4),
    # 不要忘记修改 num_classes
    bbox_head=dict(dict(num_classes=num_classes)))

# 数据集不同，dataset 输入参数也不一样
train_dataloader = dict(
    batch_size=train_batch_size_per_gpu,
    num_workers=train_num_workers,
    pin_memory=False,
    dataset=dict(
        data_root=data_root,
        metainfo=metainfo,
        ann_file='balloon/train/annotation_coco.json',
        data_prefix=dict(img='balloon/train')))

val_dataloader = dict(
    batch_size=val_batch_size_per_gpu,
    num_workers=val_num_workers,
    dataset=dict(
        metainfo=metainfo,
        data_root=data_root,
        ann_file='balloon/train/annotation_coco.json',
        data_prefix=dict(img='balloon/val/')))

test_dataloader = val_dataloader

# 默认的学习率调度器是 warmup 1000，但是 cat 数据集太小了，需要修改 为 30 iter
param_scheduler = [
    dict(
        type='LinearLR',
        start_factor=1.0e-5,
        by_epoch=False,
        begin=0,
        end=30),
    dict(
        type='CosineAnnealingLR',
        eta_min=base_lr * 0.05,
        begin=max_epochs // 2,  # max_epoch 也改变了
        end=max_epochs,
        T_max=max_epochs // 2,
        by_epoch=True,
        convert_to_iter_based=True),
]
optim_wrapper = dict(optimizer=dict(lr=base_lr))

# 第二 stage 切换 pipeline 的 epoch 时刻也改变了
_base_.custom_hooks[1].switch_epoch = max_epochs - num_epochs_stage2

val_evaluator = dict(ann_file=data_root + 'annotations/test.json')
test_evaluator = val_evaluator

# 一些打印设置修改
default_hooks = dict(
    checkpoint=dict(interval=10, max_keep_ckpts=2, save_best='auto'),  # 同时保存最好性能权重
    logger=dict(type='LoggerHook', interval=5))
train_cfg = dict(max_epochs=max_epochs, val_interval=10)

训练

训练前的可视化验证。代码如下：

from mmdet.registry import DATASETS, VISUALIZERS
from mmengine.config import Config
from mmengine.registry import init_default_scope
import matplotlib.pyplot as plt
import os.path as osp
cfg = Config.fromfile('rtmdet_tiny_1xb12-40e_cat.py')

init_default_scope(cfg.get('default_scope', 'mmdet'))

dataset = DATASETS.build(cfg.train_dataloader.dataset)
visualizer = VISUALIZERS.build(cfg.visualizer)
visualizer.dataset_meta = dataset.metainfo

plt.figure(figsize=(16, 5))

# 只可视化前 8 张图片
for i in range(8):
   item=dataset[i]

   img = item['inputs'].permute(1, 2, 0).numpy()
   data_sample = item['data_samples'].numpy()
   gt_instances = data_sample.gt_instances
   img_path = osp.basename(item['data_samples'].img_path)

   gt_bboxes = gt_instances.get('bboxes', None)
   gt_instances.bboxes = gt_bboxes.tensor
   data_sample.gt_instances = gt_instances

   visualizer.add_datasample(
            osp.basename(img_path),
            img,
            data_sample,
            draw_pred=False,
            show=False)
   drawed_image=visualizer.get_image()

   plt.subplot(2, 4, i+1)
   plt.imshow(drawed_image[..., [2, 1, 0]])
   plt.title(f"{osp.basename(img_path)}")
   plt.xticks([])
   plt.yticks([])

plt.tight_layout()
plt.show()

运行：

 python tools/train.py rtmdet_tiny_1xb12-40e_cat.py

结果：

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.741
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.846
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.823
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.496
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.854
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.228
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.784
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.818
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.733
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.892
06/10 05:35:01 - mmengine - INFO - bbox_mAP_copypaste: 0.741 0.846 0.823 0.000 0.496 0.854
06/10 05:35:01 - mmengine - INFO - Epoch(val) [100][13/13]  coco/bbox_mAP: 0.7410  coco/bbox_mAP_50: 0.8460  coco/bbox_mAP_75: 0.8230  coco/bbox_mAP_s: 0.0000  coco/bbox_mAP_m: 0.4960  coco/bbox_mAP_l: 0.8540  data_time: 0.0033  time: 0.0201

测试

推理代码：

python tools/test.py rtmdet_tiny_1xb12-40e_cat.py work_dirs/rtmdet_tiny_1xb12-40e_cat/best_coco/bbox_mAP_epoch_90.pth

测试结果：

DONE (t=0.01s).
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.745
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.837
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.815
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.472
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.870
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.230
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.782
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.822
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.692
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.911
06/10 14:19:04 - mmengine - INFO - bbox_mAP_copypaste: 0.745 0.837 0.815 0.000 0.472 0.870
06/10 14:19:04 - mmengine - INFO - Epoch(test) [13/13]  coco/bbox_mAP: 0.7450  coco/bbox_mAP_50: 0.8370  coco/bbox_mAP_75: 0.8150  coco/bbox_mAP_s: 0.0000  coco/bbox_mAP_m: 0.4720  coco/bbox_mAP_l: 0.8700  data_time: 0.5673  time: 1.1063

测试单张图片

python demo/image_demo.py 975.jpg rtmdet_tiny_1xb12-40e_cat.py --weights work_dirs/rtmdet_tiny_1xb12-40e_cat/best_coco/bbox_mAP_epoch_90.pth

测试结果：

特征图可视化

安装mmyolo，执行命令：

git clone -b tutorials https://github.com/open-mmlab/mmyolo.git 
cd mmyolo
pip install -e .

首先，resize图片，代码如下：

import cv2

img = cv2.imread('../mmdetection-tutorials/975.jpg')
h,w=img.shape[:2]
resized_img = cv2.resize(img, (640, 640))
cv2.imwrite('resized_image.jpg', resized_img)

然后再mmyolo的命令行中执行可视化，命令如下：

python demo/featmap_vis_demo.py resized_image.jpg ../mmdetection-tutorials/rtmdet_tiny_1xb12-40e_cat.py ../mmdetection-tutorials/work_dirs/rtmdet_tiny_1xb12-40e_cat/best_coco/bbox_mAP_epoch_90.pth  --target-layers backbone  --channel-reduction squeeze_mean

Grad-Based CAM 可视化

先安装Grad-Based CAM库，执行命令：

pip install grad-cam

然后再mmyolo，执行命令：

python demo/boxam_vis_demo.py resized_image.jpg ../mmdetection-tutorials/rtmdet_tiny_1xb12-40e_cat.py ../mmdetection-tutorials/work_dirs/rtmdet_tiny_1xb12-40e_cat/best_coco/bbox_mAP_epoch_90.pth  --target-layers  neck.out_convs[2]