Impressive Results when changing conf-thres and iou-thres

[pourmand1376]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

I have actually searched yolov5 and read theses related issues:

• conf_thres increased and then recall increases, data problem or others ? #7864
• Conf_thres and iou_thres  #8615
• about conf_thres issues #8302
• Result mismatch in coco metrics #6953
• Understanding F1 Curve over Confidence values #8646

Still, with my dataset (which is medical image) I get impressive results when increasing conf-thres. I am not sure if my model is doing good or actually changing conf-thres is basically wrong?

Also, PR curves for each conf-thres is different. Is this normal?
I am worried about the results because of warning that you issue when conf-thres is more than 0.001.

[Default] Conf-thres = 0.001

Conf-thres = 0.5

Conf-thres = 0.7

Conf-thres = 0.8

Conf-thres = 0.85

Conf-thres = 0.86

Conf-thres = 0.88

Conf-thres = 0.9

Additional

No response

[Zephyr69]

If you mean the conf-thres in val.py, yes it's wrong to change that. The reason is exactly what was said by the warning.

[pourmand1376]

In my specific application, I only care about precision/recall.
Is that also wrong? If yes, why would that be the case?

Also, see this comment which says that changing mAP would produce inaccurate mAP but nothing about other metrics.

[pourmand1376]

As I can see, this is the only place where conf-thres is used in val.py.

yolov5/val.py

Line 217 in 4c1784b

out = non_max_suppression(out, conf_thres, iou_thres, labels=lb, multi_label=True, agnostic=single_cls)

So, this is basically filtering out the results. Results which have lower conf-thres than defined are just ignored. This is fine by any standard.

This is where AP for each class is calculated.

yolov5/val.py

Lines 263 to 271 in 4c1784b

	# Compute metrics
	stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)] # to numpy
	if len(stats) and stats[0].any():
	tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names)
	ap50, ap = ap[:, 0], ap.mean(1) # AP@0.5, AP@0.5:0.95
	mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
	nt = np.bincount(stats[3].astype(int), minlength=nc) # number of targets per class
	else:
	nt = torch.zeros(1)

And this is the actual code for calculating average precision.

yolov5/utils/metrics.py

Lines 29 to 93 in 4c1784b

	def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=(), eps=1e-16):
	""" Compute the average precision, given the recall and precision curves.
	Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.
	# Arguments
	tp: True positives (nparray, nx1 or nx10).
	conf: Objectness value from 0-1 (nparray).
	pred_cls: Predicted object classes (nparray).
	target_cls: True object classes (nparray).
	plot: Plot precision-recall curve at mAP@0.5
	save_dir: Plot save directory
	# Returns
	The average precision as computed in py-faster-rcnn.
	"""
	# Sort by objectness
	i = np.argsort(-conf)
	tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]
	# Find unique classes
	unique_classes, nt = np.unique(target_cls, return_counts=True)
	nc = unique_classes.shape[0] # number of classes, number of detections
	# Create Precision-Recall curve and compute AP for each class
	px, py = np.linspace(0, 1, 1000), [] # for plotting
	ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000))
	for ci, c in enumerate(unique_classes):
	i = pred_cls == c
	n_l = nt[ci] # number of labels
	n_p = i.sum() # number of predictions
	if n_p == 0 or n_l == 0:
	continue
	# Accumulate FPs and TPs
	fpc = (1 - tp[i]).cumsum(0)
	tpc = tp[i].cumsum(0)
	# Recall
	recall = tpc / (n_l + eps) # recall curve
	r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases
	# Precision
	precision = tpc / (tpc + fpc) # precision curve
	p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score
	# AP from recall-precision curve
	for j in range(tp.shape[1]):
	ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j])
	if plot and j == 0:
	py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5
	# Compute F1 (harmonic mean of precision and recall)
	f1 = 2 * p * r / (p + r + eps)
	names = [v for k, v in names.items() if k in unique_classes] # list: only classes that have data
	names = dict(enumerate(names)) # to dict
	if plot:
	plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names)
	plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1')
	plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision')
	plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall')
	i = smooth(f1.mean(0), 0.1).argmax() # max F1 index
	p, r, f1 = p[:, i], r[:, i], f1[:, i]
	tp = (r * nt).round() # true positives
	fp = (tp / (p + eps) - tp).round() # false positives
	return tp, fp, p, r, f1, ap, unique_classes.astype(int)

which calls this function:

yolov5/utils/metrics.py

Lines 96 to 121 in 4c1784b

	def compute_ap(recall, precision):
	""" Compute the average precision, given the recall and precision curves
	# Arguments
	recall: The recall curve (list)
	precision: The precision curve (list)
	# Returns
	Average precision, precision curve, recall curve
	"""
	# Append sentinel values to beginning and end
	mrec = np.concatenate(([0.0], recall, [1.0]))
	mpre = np.concatenate(([1.0], precision, [0.0]))
	# Compute the precision envelope
	mpre = np.flip(np.maximum.accumulate(np.flip(mpre)))
	# Integrate area under curve
	method = 'interp' # methods: 'continuous', 'interp'
	if method == 'interp':
	x = np.linspace(0, 1, 101) # 101-point interp (COCO)
	ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate
	else: # 'continuous'
	i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes
	ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve
	return ap, mpre, mrec

Just the first stage is affected by conf-thres, so after changing this hyperparameter, the results of all metrics should be OK.

Note that ap_per_class function doesn't depend on any external thresholds. So, PR curve would not combine results for all confidence thresholds but just for one conf-threshold. This is what is see from code.

Am I missing something?

[GavinYang5]

i have the same question

[pourmand1376]

I think this may be a bug. I checked that results for recall, precision and mAP are actually wrong!

Recall should decrease as we increase conf-thres value.

[pourmand1376]

I think the bug is here. In line 217, we are filtering out based on non_max_suppression. Then we enumerate on filtered result and use filtered targets to calculate mAP. This should not be the case. We should always return unfiltered targets.

yolov5/val.py

Lines 214 to 247 in 1c5e92a

	targets[:, 2:] *= torch.tensor((width, height, width, height), device=device) # to pixels
	lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling
	t3 = time_sync()
	out = non_max_suppression(out, conf_thres, iou_thres, labels=lb, multi_label=True, agnostic=single_cls)
	dt[2] += time_sync() - t3
	# Metrics
	for si, pred in enumerate(out):
	labels = targets[targets[:, 0] == si, 1:]
	nl, npr = labels.shape[0], pred.shape[0] # number of labels, predictions
	path, shape = Path(paths[si]), shapes[si][0]
	correct = torch.zeros(npr, niou, dtype=torch.bool, device=device) # init
	seen += 1
	if npr == 0:
	if nl:
	stats.append((correct, *torch.zeros((2, 0), device=device), labels[:, 0]))
	continue
	# Predictions
	if single_cls:
	pred[:, 5] = 0
	predn = pred.clone()
	scale_coords(im[si].shape[1:], predn[:, :4], shape, shapes[si][1]) # native-space pred
	# Evaluate
	if nl:
	tbox = xywh2xyxy(labels[:, 1:5]) # target boxes
	scale_coords(im[si].shape[1:], tbox, shape, shapes[si][1]) # native-space labels
	labelsn = torch.cat((labels[:, 0:1], tbox), 1) # native-space labels
	correct = process_batch(predn, labelsn, iouv)
	if plots:
	confusion_matrix.process_batch(predn, labelsn)
	stats.append((correct, pred[:, 4], pred[:, 5], labels[:, 0])) # (correct, conf, pcls, tcls)

I've also printed target_cls.shape when passing the input to ap_per_class function like this:

def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=(), eps=1e-16):
    """ Compute the average precision, given the recall and precision curves.
    Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.
    # Arguments
        tp:  True positives (nparray, nx1 or nx10).
        conf:  Objectness value from 0-1 (nparray).
        pred_cls:  Predicted object classes (nparray).
        target_cls:  True object classes (nparray).
        plot:  Plot precision-recall curve at mAP@0.5
        save_dir:  Plot save directory
    # Returns
        The average precision as computed in py-faster-rcnn.
    """
    print(f"TP Shape: {tp.shape}")
    print(f"Conf Shape : {conf.shape}")
    print(f"predicted_cls: {pred_cls.shape}")
    print(f"target_cls: {target_cls.shape}")

## using conf-thres 0.85
TP Shape: (33, 10)
Conf Shape : (33,)
predicted_cls: (33,)
target_cls: (30,)
#  -------------
# using conf-thres 0.001
TP Shape: (25239, 10)
Conf Shape : (25239,)
predicted_cls: (25239,)
target_cls: (281,)
# -------------
# using conf-thres 0.001
TP Shape: (24777, 10)
Conf Shape : (24777,)
predicted_cls: (24777,)
target_cls: (266,)
# --------------

Even when having same conf-thres as recommended, we end up calculating the result wrong! Target_cls should always be the same, no matter what conf-thres is, this is because number of target labels per class is always the same.

This way, even setting conf-thres to 0.001 gives wrong results.

If you look at labels count column in every threshold above, you understand that this is wrong. Labels count shouldn't change for each run!

I will do a PR soon.

[pourmand1376]

Now with PR #8686, my results are consistent and everything is fine. You can also see that label's count is persistent in each run and it is equal to actual labels that I have.

However, my model is not doing any good, so I can not publish the results! Before that, I was actually happy about the result.

Here's the results after fixing the bug:

Conf-thres = 0.001

Conf-thres = 0.2

Conf-thres = 0.5

Conf-thres = 0.7

Conf-thres = 0.85

Conf-thres = 0.9

[pourmand1376]

We can not just say that changing conf-thres is wrong.

See this paper which has demonstrated the importance of confidence threshold. It is interesting that it has actually done some experiments using Yolov5.

Confidence Score: The Forgotten Dimension of Object Detection Performance Evaluation

Page 15 of the paper:

[pourmand1376]

This PR is not required actually. If you have this problem, you should just update your forked repo to latest version. This has been fixed a little earlier.

[kaminetzky]

I was having the same issues before updating my repo. Thank you for your analysis!

[glenn-jocher]

@kaminetzky you're welcome! It's great to hear that your issues have been addressed. Feel free to reach out if you need further assistance. Good luck with your work!