TOOD

TOOD copied to clipboard

Hi, thanks for your wonderful work, I have a question how do you plot the detection result in Figure1? especially the prediction score map and localization map.

You can plot and save the figure with matplotlib, for example:

import matplotlib.pyplot as plt

fig = plt.figure(figsize=(18,20)) ax = fig.add_subplot(111) im = ax.imshow(score_map, cmap=plt.cm.jet) cb = plt.colorbar(im, fraction=0.058) cb.ax.tick_params(labelsize=64) plt.axis('off') plt.savefig('demo.jpg')

You can plot and save the figure with matplotlib, for example:

import matplotlib.pyplot as plt

fig = plt.figure(figsize=(18,20)) ax = fig.add_subplot(111) im = ax.imshow(score_map, cmap=plt.cm.jet) cb = plt.colorbar(im, fraction=0.058) cb.ax.tick_params(labelsize=64) plt.axis('off') plt.savefig('demo.jpg')

How do I generate score_map to plot?

@iumyx2612 score_map is the score map predicted from the classifier.

@iumyx2612 score_map is the score map predicted from the classifier.

I was able to produce Classification score map, but how do you get the IoU score map? There's no IoU branch in TOOD

@iumyx2612 score_map is the score map predicted from the classifier.

I was able to produce Classification score map, but how do you get the IoU score map? There's no IoU branch in TOOD

The IoU score map is computed between the predicted bounding boxes and the GT bounding box.

@iumyx2612 score_map is the score map predicted from the classifier.

I was able to produce Classification score map, but how do you get the IoU score map? There's no IoU branch in TOOD

The IoU score map is computed between the predicted bounding boxes and the GT bounding box.

Thank you so much!!!

@iumyx2612 score_map is the score map predicted from the classifier.

I was able to produce Classification score map, but how do you get the IoU score map? There's no IoU branch in TOOD

The IoU score map is computed between the predicted bounding boxes and the GT bounding box.

Thank you so much!!!

Did you produce the plot result? Can you share me a link?

You can plot and save the figure with matplotlib, for example:

import matplotlib.pyplot as plt

fig = plt.figure(figsize=(18,20)) ax = fig.add_subplot(111) im = ax.imshow(score_map, cmap=plt.cm.jet) cb = plt.colorbar(im, fraction=0.058) cb.ax.tick_params(labelsize=64) plt.axis('off') plt.savefig('demo.jpg')

I want to make a comparison with my model. Can you tell me the name of this picture？

@iumyx2612 score_map is the score map predicted from the classifier.

I was able to produce Classification score map, but how do you get the IoU score map? There's no IoU branch in TOOD

The IoU score map is computed between the predicted bounding boxes and the GT bounding box.

Thank you so much!!!

Did you produce the plot result? Can you share me a link?

Sorry I didn't upload it to anywhere, here's my code

def analyze_aligment(config_file,
                     checkpoint_file,
                     category: int=0,
                     img_id: int=0):
    """ Script to visualize alignment for classification and regression
    Args:
        config_file: model config file
        checkpoint_file: model checkpoint
        category: class to visualize
        img_id: image index from dataset
    Example:
        >>> config_file = "configs/tood/tood_r50_voc.py"
        >>> checkpoint_file = "ckpts/tood_r50_fpn_1x_coco.pth"
        >>> analyze_aligment(config_file, checkpoint_file)
    """
    iou_calculator = BboxOverlaps2D()
    cfg = Config.fromfile(config_file)
    device = get_device()

    model = init_detector(config_file, checkpoint_file, device=device)
    model.eval()
    model_head = model.bbox_head

    if cfg.data.train['type'] != 'MultiImageMixDataset':
        val_dataset = copy.deepcopy(cfg.data.train)
    else:
        val_dataset = copy.deepcopy(cfg.data.test)
    dataset = build_dataset(val_dataset)

    img = dataset[img_id]['img'].data.unsqueeze(0).to(device)
    gt_bboxes = dataset[img_id]['gt_bboxes'].data.to(device)

    with torch.no_grad():
        backbone_feat = model.extract_feat(img) # tuple, len = num_level
        cls_scores, bbox_preds = model_head(backbone_feat)
        # cls_score (batch, classes, H, W)
        # bbox_pred (batch, 4, H, W)

        flatten_bbox_preds = []
        for bbox_pred, stride in zip(bbox_preds, model_head.prior_generator.strides):
            flatten_bbox_pred = bbox_pred.permute(0, 2, 3, 1).reshape(1, -1, 4) * stride[0]
            flatten_bbox_preds.append(flatten_bbox_pred)

        iou_scores = []
        for flatten_bbox_pred, bbox_pred in zip(flatten_bbox_preds, bbox_preds):
            flatten_bbox_pred = flatten_bbox_pred.squeeze() # (N, H*W, 4) -> (H*W, 4)
            overlaps = iou_calculator(flatten_bbox_pred, gt_bboxes)  # (H*W, num_gts)
            orig_shape = bbox_pred.shape[2:]
            overlaps = overlaps.reshape((orig_shape[0], orig_shape[1], overlaps.shape[-1])) # (H, W, num_gts)
            iou_scores.append(overlaps)

        cls_scores = [cls_score.squeeze().permute(1, 2, 0) for cls_score in cls_scores]

    img = img.squeeze().permute(1, 2, 0).cpu().numpy()

    for scale, (cls_score, iou_score) in enumerate(zip(cls_scores, iou_scores)):
        cls_score = cls_score[:, :, category] # (H, W, 1)

        cls_score = cls_score.cpu().numpy()
        iou_score = iou_score.cpu().numpy()
        num_bboxes = iou_score.shape[-1]

        fig = plt.figure(figsize=(10, 10))
        fig.suptitle(f"Scale {scale + 1}")

        ax1 = fig.add_subplot(num_bboxes, 2, 1)
        ax2 = fig.add_subplot(num_bboxes, 2, 2)
        ax1.set_title("Cls Score")
        ax2.set_title("Image")
        im1 = ax1.imshow(cls_score, cmap=plt.cm.jet)
        im2 = ax2.imshow(img)
        cb1 = plt.colorbar(im1, fraction=0.05, ax=ax1)
        cb1.ax.tick_params(labelsize=5)

        for i in range(num_bboxes):
            score = iou_score[:, :, i]
            ax = fig.add_subplot(num_bboxes, 2, i + 3)
            ax.set_title(f"IoU Score {i + 1}")
            im = ax.imshow(score, cmap=plt.cm.jet)
            cb = plt.colorbar(im, fraction=0.05, ax=ax)
            cb.ax.tick_params(labelsize=5)

        plt.show()

How can we visualise heatmaps??? for instance in ceneternet i want to visualize the heatmaps of its output