Shulou(Shulou.com)05/31 Report--

This article mainly introduces the relevant knowledge of "anchor setting method of YOLOv5 target detection". The editor shows you the operation process through an actual case. The operation method is simple, fast and practical. I hope this article "anchor setting method of YOLOv5 target detection" can help you solve the problem.

Preface

Yolo algorithm, as a leader in the field of one-stage, uses the anchor-based method for target detection, uses different scales of anchor to directly regression the target box and outputs the location and category confidence of the target box at one time.

Why use anchor for detection?

The initial training process of the initial YOLOv1 is very unstable. During the design of the YOLOv2, the author observed a large number of pictures of ground truth and found that the target instances of the same category have similar gt aspect ratio: for example, cars, gt are short and fat rectangles; for example, pedestrians, gt are tall and thin rectangles. So inspired by this, the author prepares several bounding box with high probability in advance from the data set, and then uses them as a benchmark for prediction.

The detection process of anchor

First of all, the input image size of the coco dataset used in yolov5 is 640x640, but the input size of the training process is not unique, because v5 can use masaic enhancement technology to make the parts of the four pictures into a certain size input picture. However, if you need to use pre-training weights, it is best to adjust the input image size to the same size as the author, and the input image size must be a multiple of 32, which is related to the following anchor detection stage.

The above picture is the network structure diagram of v5 v6.0 drawn by myself. When our input size is 640mm / 640, we will get three different scales of output: 80x80 (640cm 8), 40x40 (640max 16), and 20x20 (640max 32), that is, the output of the CSP2_3 module in the image above.

Anchors:-[10 recollection 13, 16 recollection 30, 3 3jue 23] # P3 Grammer 8-[30 recorder 61,62 pyrrine 45, 59119] # P4 Placement 16-[116 record90, 156 recovery198, 373326] # P5and32

Among them, 80x80 represents the shallow feature map (P3), which contains more low-level information, so it is suitable for detecting small targets, so the anchor scale used in this feature map is smaller; similarly, 20x20 represents the deep feature map (P5), which contains more high-level information, such as outline, structure and other information, so it is suitable for large target detection, so the anchor scale used in this feature map is larger. On the other 40x40 feature map (P4), an anchor between these two scales is used to detect medium-sized targets. The reason why yolov5 can detect cross-scale targets efficiently and quickly is due to the idea of using different scales of anchor for different feature maps.

The above is the specific explanation of anchors in yolov5.

Anchor production process

For most images, because their size does not match our preset input size, resize is done in the input phase, resulting in a change in the pre-marked bounding box size. The anchor is calculated according to the size of the bounding box we input into the network, so there is a process of anchor re-clustering in this resize process. Under the yolov5/utils/autoanchor.py file, there is a function kmeans_anchor, which calculates anchor by the method of kmeans. The details are as follows:

Def kmean_anchors (dataset='./data/coco128.yaml', nasty 9, img_size=640, thr=4.0, gen=1000, verbose=True): "Creates kmeans-evolved anchors from training dataset Arguments: dataset: path to data.yaml, or a loaded dataset n: number of anchors img_size: image size used for training thr: anchor-label wh ratio threshold hyperparameter hyp ['anchor_t'] used for training Default=4.0 gen: generations to evolve anchors using genetic algorithm verbose: print all results Return: K: kmeans evolved anchors Usage: from utils.autoanchor import * _ = kmean_anchors () "from scipy.cluster.vq import kmeans thr = 1. / thr prefix = colorstr ('autoanchor:') def metric (k, wh): # compute metrics r = wh [:, None] / k [None] x = torch.min (r, 1. / r) .min (2) [0] # ratio metric # x = wh_iou (wh Torch.tensor (k)) # iou metric return x, x.max (1) [0] # x, best_x def anchor_fitness (k): # mutation fitness _, best = metric (torch.tensor (k, dtype=torch.float32)) Wh) return (best * (best > thr). Float (). Mean () # fitness def print_results (k): K = k [np.argsort (k.prod (1))] # sort small to large x, best = metric (k, wh0) bpr, aat = (best > thr). Float (). Mean (), (x > thr). Float (). Mean () * n # best possible recall Anch > thr print (f'{prefix} thr= {thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr') print (f'{prefix} n = {n}, img_size= {img_size}, metric_all= {x.mean (): .3f} / {best.mean (): .3f}-mean/best) X in enumerate (k): print ('% I thr% i'% (round (x [0]), round (x [1])), end=','if I

< len(k) - 1 else '\n') # use in *.cfg return k if isinstance(dataset, str): # *.yaml file with open(dataset, errors='ignore') as f: data_dict = yaml.safe_load(f) # model dict from datasets import LoadImagesAndLabels dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True) # Get label wh shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True) wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh # Filter i = (wh0 < 3.0).any(1).sum() if i: print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.') wh = wh0[(wh0 >

.any (1)] # filter > 2 pixels # wh = wh * (np.random.rand (wh.shape [0], 1) * 0.9 + 0.1) # multiply by random scale 0-1 # Kmeans calculation print (f'{prefix} Running kmeans for {n} anchors on {len (wh)} points...') S = wh.std (0) # sigmas for whitening k, dist = kmeans (wh / s, n, iter=30) # points, mean distance assert len (k) = = n, f'{prefix} ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len (k)}'k * = s wh = torch.tensor (wh, dtype=torch.float32) # filtered wh0 = torch.tensor (wh0, dtype=torch.float32) # unfiltered k = print_results (k) # Plot # k D = [None] * 20, [None] * 20 # for i in tqdm (range (1,21)): # k [I-1], d [I-1] = kmeans (wh / s, I) # points, mean distance # fig, ax = plt.subplots (1,2, figsize= (14,7), tight_layout=True) # ax = ax.ravel () # ax [0] .plot (np.arange (1,21), np.array (d) * * 2 Marker='.') # fig, ax = plt.subplots (1,2, figsize= (14,7)) # plot wh # ax [0] .hist (wh [wh [:, 0])

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.