AI Era Survival: Using YOLOv3 for Accurate Pig Detection

YOLOv3 Core Improvements

Backbone upgrade : Darknet‑53 (53 convolutional layers with residual blocks) replaces Darknet‑19, providing deeper feature extraction and smoother gradient flow.

Three‑scale feature fusion : Inspired by FPN, YOLOv3 outputs 13×13, 26×26 and 52×52 feature maps. The 52×52 map preserves fine details for tiny piglets, the 26×26 map handles medium‑sized pigs, and the 13×13 map captures large adult pigs.

Anchor refinement : Each scale receives three anchors (nine total) generated by separate K‑means clustering, better matching target sizes at each scale.

Classification head : Softmax is replaced by a sigmoid + binary cross‑entropy loss, enabling multi‑label classification such as “pig + marked pig”.

Dynamic input resizing : Input size is randomly chosen among multiples of 32 (e.g., 320, 416, 608) instead of a fixed 416×416, improving generalisation to different camera distances and pen sizes.

Pig Detection Model

Dataset: 2,456 pig images split into 1,720 training, 491 validation and 245 test samples.

class Yolov3(nn.Module):
    def __init__(self, cfg, is_val=False) -> None:
        super(Yolov3, self).__init__()
        self.cfg = cfg
        self.num_classes = cfg.num_classes
        self.topk_candidates = cfg.val_topk if is_val else cfg.test_topk
        self.conf_thresh = cfg.val_conf_thresh if is_val else cfg.test_conf_thresh
        self.nms_thresh = cfg.val_nms_thresh if is_val else cfg.test_nms_thresh
        self.no_multi_labels = False if is_val else True
        # Backbone
        self.backbone = Yolov3Backbone(cfg)
        self.pyramid_feat_dims = self.backbone.feat_dims[-3:]
        # Neck: SPPF
        self.neck = SPPF(cfg, self.pyramid_feat_dims[-1], self.pyramid_feat_dims[-1])
        self.pyramid_feat_dims[-1] = self.neck.out_dim
        # Neck: FPN
        self.fpn = Yolov3FPN(cfg, self.pyramid_feat_dims)
        # Head
        self.head = Yolov3DetHead(cfg, self.fpn.out_dims)
        # Prediction layer
        self.pred = Yolov3DetPredLayer(cfg)

Training command:

nohup python -u train.py --cuda --dataset custom --root dataset/MonitorPigDataset --model yolov3_s --batch_size 16 >> train.log 2>&1 &

Training Results

After 300 epochs the model converged. Validation metrics (averaged) are:

Averaged stats: lr: 0.000010 size: 640 gnorm: 5.5 loss_obj: 1.4137 (1.4709) loss_cls: 0.3966 (0.3938) loss_box: 0.1408 (0.1403) losses: 0.6211 (0.6415)
AP@[0.50:0.95] = 0.673
[email protected] = 0.981
[email protected] = 0.829
AP small = 0.021
AP medium = 0.655
AP large = 0.686
AR@[0.50:0.95] (maxDets=100) = 0.750
AR small = 0.333
AR medium = 0.749
AR large = 0.751

Medium and large pigs achieve AP above 0.65, while small piglets obtain AP of 0.021. Compared with a YOLOv2 baseline the three‑scale fusion improves detection of small targets by more than 20 %.

Conclusion

The deeper Darknet‑53 backbone, multi‑scale feature maps, refined anchors and sigmoid‑based classification together make YOLOv3 well‑suited for animal detection tasks. With a modest dataset and the described training strategy, high‑precision pig detection is achievable without resorting to custom heavyweight networks.