Skip to content
DeepForgeby svx
Advancedcomputer-visioncnndetectionsegmentationvision-transformers

Computer Vision Deep Learning

30 problems·~12h·4 stages

The deep-learning side of vision, from convolution arithmetic to ViTs and depth. You will count the FLOPs and parameters of modern conv blocks, fold batch norm, schedule augmentations, match anchors, score detections and segmentations, and patchify images for transformers. Take it after the vision starter and the deep-learning essentials.

Your progress

0/30 · 0%

Path readiness 0%Easy 0% · Medium 0% · Hard 0%

Continue path · Conv Output Size

What you will learn

  • Compute convolution output shapes, parameter counts, and FLOPs at scale
  • Fold normalisation into convolutions and schedule modern augmentations
  • Match anchors, score detections with mAP, and evaluate segmentation masks
  • Patchify images for ViTs and recover depth from stereo disparity

Before you start

1Convolution at Scale

0/8

Convolution is a shape and budget calculation before it is code. Work out output sizes for dilated, depthwise, group, and 3D convolutions, count parameters and FLOPs, and fuse conv with ReLU.

Stage checkpoint

0/2 to pass

A checkpoint is a short retrieval test drawn from this stage: solve 2 of 3 from memory to mark the stage complete. The stage stays open either way.

Tried this set and did not pass? Record it — the next set swaps in fresh problems from the same stage.
ProjectRecommended next stepBuild a CNN from ScratchHard

2Normalisation and Augmentation

0/7

Keep deep vision models stable and their training data varied: batch norm at inference, folding batch norm into the preceding convolution, and the random-choice and magnitude schedules behind RandAugment and TrivialAugment, plus a seeded random resized crop.

Stage checkpoint

0/2 to pass

A checkpoint is a short retrieval test drawn from this stage: solve 2 of 3 from memory to mark the stage complete. The stage stays open either way.

Tried this set and did not pass? Record it — the next set swaps in fresh problems from the same stage.
ProjectRecommended next stepBuild a CNN from ScratchHard

3Detection and Segmentation

0/8

Teach a model where things are: generate and match anchor boxes, pool regions of interest, build pairwise IoU matrices, compute mAP at an IoU threshold, and score segmentation masks with Hamming loss and class-weighted IoU.

Stage checkpoint

0/2 to pass

A checkpoint is a short retrieval test drawn from this stage: solve 2 of 3 from memory to mark the stage complete. The stage stays open either way.

Tried this set and did not pass? Record it — the next set swaps in fresh problems from the same stage.
ProjectRecommended next stepBuild a CNN from ScratchHard

4Vision Transformers and 3D

0/7

Beyond convolution: patchify images and embed patches for a ViT, merge patches to downsample, unpatchify back to pixels, measure masked-patch reconstruction loss, and recover depth from stereo disparity with scale-invariant error.

Stage checkpoint

0/2 to pass

A checkpoint is a short retrieval test drawn from this stage: solve 2 of 3 from memory to mark the stage complete. The stage stays open either way.

Tried this set and did not pass? Record it — the next set swaps in fresh problems from the same stage.
ProjectRecommended next stepBuild a CNN from ScratchHard

Capstone

Project

Close the path with one real artifact from the catalog.

Build a CNN from Scratch

Assemble convolution, pooling, and softmax into a classifier you can trace end to end.