how can I get the average precision recall curve for multiclass? about object-detection-metrics

[rafaelpadilla]

Dear @hwjung92,

If you want to show the plots with the Average Precision graphs for each class, you need to call the function evaluator.PlotPrecisionRecallCurve and set the parameter showInterpolatedPrecision=True.

I encourage you to see this sample.

If you want to use the command line, you can use --savepath and omit the --noplot argument.

Please, close this issue if your question regarding this issue was answered properly.

Regards,
Rafael

[hwjung92]

Thank you for rapid response.
My question is how to draw average precision graph for all calsses, instead of each class.
Actually, I modified simply GetPascalVOCMetrics() function like below.
I summed all FP and TP of each class, and calculate acc_FP and acc_TP.

        FPS.append(FP)
        TPS.append(TP)

    # compute precision, recall and average precision
    FPS2 = [j for sub in FPS for j in sub]
    TPS2 = [j for sub in TPS for j in sub]   
    acc_FP = np.cumsum(FPS2)
    acc_TP = np.cumsum(TPS2)

It works, Thank you.

[hwjung92]

I want to draw the graph about Average Precision-Recall of all classes.
My implemented function is posted like below.

    def GetPascalVOCMetricsForAllClass(self,
                        boundingboxes,
                        IOUThreshold=0.5,
                        method=MethodAveragePrecision.EveryPointInterpolation):
    ret = []  # list containing metrics (precision, recall, average precision) of each class
    # List with all ground truths (Ex: [imageName,class,confidence=1, (bb coordinates XYX2Y2)])
    groundTruths = []
    # List with all detections (Ex: [imageName,class,confidence,(bb coordinates XYX2Y2)])
    detections = []
    # Get all classes
    classes = []
    # Loop through all bounding boxes and separate them into GTs and detections
    for bb in boundingboxes.getBoundingBoxes():
        # [imageName, class, confidence, (bb coordinates XYX2Y2)]
        if bb.getBBType() == BBType.GroundTruth:
            groundTruths.append([
                bb.getImageName(),
                bb.getClassId(), 1,
                bb.getAbsoluteBoundingBox(BBFormat.XYX2Y2)
            ])
        else:
            detections.append([
                bb.getImageName(),
                bb.getClassId(),
                bb.getConfidence(),
                bb.getAbsoluteBoundingBox(BBFormat.XYX2Y2)
            ])
        # get class
        if bb.getClassId() not in classes:
            classes.append(bb.getClassId())
    classes = sorted(classes)
    # Precision x Recall is obtained individually by each class
    # Loop through by classes
    npos=[]
    TPS=[]
    FPS=[]
    for c in classes:
        # Get only detection of class c
        dects = []
        [dects.append(d) for d in detections if d[1] == c]
        # Get only ground truths of class c
        gts = []
        [gts.append(g) for g in groundTruths if g[1] == c]
        npos.append(len(gts))
        # sort detections by decreasing confidence
        dects = sorted(dects, key=lambda conf: conf[2], reverse=True)
        TP = np.zeros(len(dects))
        FP = np.zeros(len(dects))
        # create dictionary with amount of gts for each image
        det = Counter([cc[0] for cc in gts])
        for key, val in det.items():
            det[key] = np.zeros(val)
        # Loop through detections
        for d in range(len(dects)):
            # Find ground truth image
            gt = [gt for gt in gts if gt[0] == dects[d][0]]
            iouMax = sys.float_info.min
            for j in range(len(gt)):
                # print('Ground truth gt => %s' % (gt[j][3],))
                iou = Evaluator.iou(dects[d][3], gt[j][3])
                if iou > iouMax:
                    iouMax = iou
                    jmax = j
            # Assign detection as true positive/don't care/false positive
            if iouMax >= IOUThreshold:
                if det[dects[d][0]][jmax] == 0:
                    TP[d] = 1  # count as true positive
                    det[dects[d][0]][jmax] = 1  # flag as already 'seen'
                else:
                    FP[d] = 1  # count as false positive
            # - A detected "cat" is overlaped with a GT "cat" with IOU >= IOUThreshold.
            else:
                FP[d] = 1  # count as false positive
        FPS.append(FP)
        TPS.append(TP)
    # compute precision, recall and average precision
    FPS2 = [j for sub in FPS for j in sub]
    TPS2 = [j for sub in TPS for j in sub]    #np.array(TPS).flatten()
    #FPS2 = np.reshape(FPS, (1,np.product(FPS.shape)))
    acc_FP = np.cumsum(FPS2)
    acc_TP = np.cumsum(TPS2)
    rec = acc_TP / np.sum(npos)
    prec = np.divide(acc_TP, (acc_FP + acc_TP))
    # Depending on the method, call the right implementation
    if method == MethodAveragePrecision.EveryPointInterpolation:
        [ap, mpre, mrec, ii] = Evaluator.CalculateAveragePrecision(rec, prec)
    else:
        [ap, mpre, mrec, _] = Evaluator.ElevenPointInterpolatedAP(rec, prec)
    # add class result in the dictionary to be returned
    ret = {
        'class':'all classes',
        'precision': prec,
        'recall': rec,
        'AP': ap,
        'interpolated precision': mpre,
        'interpolated recall': mrec,
        'total positives': np.sum(npos),
        'total TP': np.sum(TPS2),
        'total FP': np.sum(FPS2)
    }
    return ret

I found that the graph of each class is correct.
However, the final image is not correct, because it is summary graph instead of average.
Is there any someone who can help me?

[rafaelpadilla]

Dear @hwjung92,

I don’t understand the need of plotting a mAP graph. The mAP is a number that evaluates your detections over all classes.

Sorry, but I don’t know what your need is.

[hwjung92]

Sorry for confusing my question. I misunderstood the concept of mAP. Actually I want to draw the average Precision-Recall graph of all class. How can I solve this issue?

[hwjung92]

I simply obtain average FP and TP of all classes, and calculate Precision and Recall.
The new result looks like average PRcure of all classes.
If there is the better way or idea, let me know.
Thank you.

[rafaelpadilla]

Now I see what you want to do. :) But I really can’t tell you if this is the correct way of doing it. Pascal VOC does have AP for all classes at the same plot. Before publishing your results, you must guarantee your metric is implemented correctly, so your results aren’t biased. That’s why I recommend you to look for other competitions and their official codes (like COCO) to verify if it is done this way. I would need to see some reference implementations or published papers to confirm if this is the right way of doing it.