org.cripac.isee.alg.pedestrian.attr

Class DeepMARCaffe2Native

java.lang.Object

org.cripac.isee.alg.pedestrian.attr.DeepMARCaffe2Native

All Implemented Interfaces:

BatchRecognizer, DeepMAR, DeepMARCaffe2, Recognizer

public class DeepMARCaffe2Native
extends Object
implements DeepMARCaffe2, BatchRecognizer

The class DeepMARCaffe2Native recognizes attributes from pedestrian images using native libraries. The recognition functions are thread safe, but cannot benefit from multi-threading. To benefit from multi-threading and multi-GPU, create multiple instances of this class in different threads, each using a different GPU.

Nested Class Summary

Nested classes/interfaces inherited from interface org.cripac.isee.alg.pedestrian.attr.DeepMAR

DeepMAR.PointerManager

Field Summary

Fields

Modifier and Type	Field and Description
private int	lastBatchSize
private Logger	logger
private float[][]	multiOutputBuf
private long	net
private float[]	outputBuf

Fields inherited from interface org.cripac.isee.alg.pedestrian.attr.DeepMAR

ATTR_LIST, INPUT_HEIGHT, INPUT_WIDTH, MEAN_PIXEL, POINTERS, random, REG_COEFF

Constructor Summary

Constructors

Constructor and Description
DeepMARCaffe2Native(String gpu, File init_net, File predict_net, Logger logger) Create an instance of DeepMARCaffe2Native.
DeepMARCaffe2Native(String gpu, Logger logger) Create an instance of DeepMARCaffe2Native.

Method Summary

Modifier and Type	Method and Description
protected void	finalize() Free the network.
private void	free(long p) Free the network.
private long	initialize(int gpu, String pbPath, String modelPath) Initialize the native DeepMAR network.
void	recognize(long net, float[][] pixelBytes, float[][] outputBuf) Recognize attributes from a batch of images using the network.
void	recognize(long net, float[] pixelBytes, float[] outputBuf) Recognize attributes from an image using the network.
Attributes	recognize(Tracklet.BoundingBox bbox) Recognize attributes from a pedestrian bounding box image.
Attributes[]	recognize(Tracklet.BoundingBox[] bboxes) Recognize attributes from a batch of pedestrian bounding box images.
Attributes	recognize(Tracklet tracklet) Recognize attributes from a pedestrian tracklet.

Methods inherited from class java.lang.Object

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

Methods inherited from interface org.cripac.isee.alg.pedestrian.attr.DeepMARCaffe2

getInitNetProtobuf, getPredictNetModel

Methods inherited from interface org.cripac.isee.alg.pedestrian.attr.DeepMAR

fillAttributes, preprocess, randomlyPickGPU

Field Detail

net

private long net

outputBuf

private final float[] outputBuf

multiOutputBuf

private float[][] multiOutputBuf

lastBatchSize

private int lastBatchSize

logger

private Logger logger

Constructor Detail

DeepMARCaffe2Native

public DeepMARCaffe2Native(String gpu,
                           @Nonnull
                           File init_net,
                           @Nonnull
                           File predict_net,
                           @Nonnull
                           Logger logger)
                    throws FileNotFoundException,
                           AccessDeniedException,
                           CharacterCodingException

Create an instance of DeepMARCaffe2Native. The protocol and weights are directly loaded from local files.

Parameters:

gpu - index of GPU to use. If multiple GPU IDs are provided (separated by comma), randomly select one.

init_net - DeepMARCaffe2Native init net protobuf file.

predict_net - DeepMARCaffe2Native predict net protobuf file.

Throws:

FileNotFoundException

AccessDeniedException

CharacterCodingException

DeepMARCaffe2Native

public DeepMARCaffe2Native(String gpu,
                           @Nonnull
                           Logger logger)
                    throws IOException

Create an instance of DeepMARCaffe2Native. The protocol and weights are retrieved from the JAR.

Parameters:

gpu - index of GPU to use. If multiple GPU IDs are provided (separated by comma), randomly select one.

Throws:

IOException

Method Detail

initialize

private long initialize(int gpu,
                        @Nonnull
                        String pbPath,
                        @Nonnull
                        String modelPath)

Initialize the native DeepMAR network.

Parameters:

gpu - index of GPU to run on.

pbPath - ASCII path of protocol buffer.

modelPath - ASCII path of Caffe model.

Returns:

pointer to the allocated native network.

free

private void free(long p)

Free the network.

Parameters:

p - the pointer to the network.

recognize

public void recognize(long net,
                      @Nonnull
                      float[] pixelBytes,
                      @Nonnull
                      float[] outputBuf)

Recognize attributes from an image using the network.

Parameters:

net - the pointer to the network.

pixelBytes - pixel bytes retrieved from the image.

outputBuf - buffer for returning the FC8 result.

recognize

public void recognize(long net,
                      @Nonnull
                      float[][] pixelBytes,
                      @Nonnull
                      float[][] outputBuf)

Recognize attributes from a batch of images using the network.

Parameters:

net - the pointer to the network.

pixelBytes - pixel bytes retrieved from each of the images in the batch.

outputBuf - buffer for returning the FC8 results for each of the images in the batch.

finalize

protected void finalize()
                 throws Throwable

Free the network.

Overrides:

finalize in class Object

Throws:

Throwable - on failure finalizing the super class.

recognize

@Nonnull
public Attributes recognize(@Nonnull
                                     Tracklet tracklet)

Recognize attributes from a pedestrian tracklet.

Specified by:

recognize in interface Recognizer

Parameters:

tracklet - a pedestrian tracklet.

Returns:

attributes of the pedestrian specified by the tracklet.

recognize

@Nonnull
public Attributes recognize(@Nonnull
                                     Tracklet.BoundingBox bbox)

Recognize attributes from a pedestrian bounding box image.

Parameters:

bbox - the bounding box containing the target pedestrian image.

Returns:

attributes of the pedestrian in the image.

recognize

@Nonnull
public Attributes[] recognize(@Nonnull
                                       Tracklet.BoundingBox[] bboxes)

Recognize attributes from a batch of pedestrian bounding box images.

Specified by:

recognize in interface BatchRecognizer

Parameters:

bboxes - a batch of bounding boxes containing the images.

Returns:

attributes of the pedestrians in the images.