org.deeplearning4j.nn.layers.objdetect

Class YoloUtils

java.lang.Object

org.deeplearning4j.nn.layers.objdetect.YoloUtils

public class YoloUtils
extends Object

Constructor Summary

Constructors

Constructor and Description
YoloUtils()

Method Summary

Modifier and Type	Method and Description
static INDArray	activate(INDArray boundingBoxPriors, INDArray input) Essentially: just apply activation functions...
static INDArray	activate(INDArray boundingBoxPriors, INDArray input, boolean nchw)
static INDArray	activate(@NonNull INDArray boundingBoxPriors, @NonNull INDArray input, boolean nchw, LayerWorkspaceMgr layerWorkspaceMgr)
static INDArray	activate(@NonNull INDArray boundingBoxPriors, @NonNull INDArray input, LayerWorkspaceMgr layerWorkspaceMgr)
static List<DetectedObject>	getPredictedObjects(INDArray boundingBoxPriors, INDArray networkOutput, double confThreshold, double nmsThreshold) Given the network output and a detection threshold (in range 0 to 1) determine the objects detected by the network. Supports minibatches - the returned DetectedObject instances have an example number index. Note that the dimensions are grid cell units - for example, with 416x416 input, 32x downsampling by the network (before getting to the Yolo2OutputLayer) we have 13x13 grid cells (each corresponding to 32 pixels in the input image).
static double	iou(DetectedObject o1, DetectedObject o2) Returns intersection over union (IOU) between o1 and o2.
static void	nms(List<DetectedObject> objects, double iouThreshold) Performs non-maximum suppression (NMS) on objects, using their IOU with threshold to match pairs.
static double	overlap(double x1, double x2, double x3, double x4) Returns overlap between lines [x1, x2] and [x3.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

YoloUtils

public YoloUtils()

Method Detail

activate

public static INDArray activate(INDArray boundingBoxPriors,
                                INDArray input)

Essentially: just apply activation functions... For NCHW format. For NCHW format, use one of the other activate methods

activate

public static INDArray activate(INDArray boundingBoxPriors,
                                INDArray input,
                                boolean nchw)

activate

public static INDArray activate(@NonNull
                                @NonNull INDArray boundingBoxPriors,
                                @NonNull
                                @NonNull INDArray input,
                                LayerWorkspaceMgr layerWorkspaceMgr)

activate

public static INDArray activate(@NonNull
                                @NonNull INDArray boundingBoxPriors,
                                @NonNull
                                @NonNull INDArray input,
                                boolean nchw,
                                LayerWorkspaceMgr layerWorkspaceMgr)

overlap

public static double overlap(double x1,
                             double x2,
                             double x3,
                             double x4)

Returns overlap between lines [x1, x2] and [x3. x4].

iou

public static double iou(DetectedObject o1,
                         DetectedObject o2)

Returns intersection over union (IOU) between o1 and o2.

nms

public static void nms(List<DetectedObject> objects,
                       double iouThreshold)

Performs non-maximum suppression (NMS) on objects, using their IOU with threshold to match pairs.

getPredictedObjects

public static List<DetectedObject> getPredictedObjects(INDArray boundingBoxPriors,
                                                       INDArray networkOutput,
                                                       double confThreshold,
                                                       double nmsThreshold)

Given the network output and a detection threshold (in range 0 to 1) determine the objects detected by the network.
Supports minibatches - the returned DetectedObject instances have an example number index.
Note that the dimensions are grid cell units - for example, with 416x416 input, 32x downsampling by the network (before getting to the Yolo2OutputLayer) we have 13x13 grid cells (each corresponding to 32 pixels in the input image). Thus, a centerX of 5.5 would be xPixels=5.5x32 = 176 pixels from left. Widths and heights are similar: in this example, a with of 13 would be the entire image (416 pixels), and a height of 6.5 would be 6.5/13 = 0.5 of the image (208 pixels).

Parameters:

boundingBoxPriors - as given to Yolo2OutputLayer

networkOutput - 4d activations out of the network

confThreshold - Detection threshold, in range 0.0 (least strict) to 1.0 (most strict). Objects are returned where predicted confidence is >= confThreshold

nmsThreshold - passed to nms(List, double) (0 == disabled) as the threshold for intersection over union (IOU)

Returns:

List of detected objects