Intermediate everything in Computer Vision, page 3
Why U-Net skip connections matter for denoising
Skips carry high-resolution spatial detail from encoder to decoder, preserving fine structure lost in downsampling and easing gradient flow, which lets the model restore detail while removing…
Evaluating generative models with FID versus IS
FID compares Inception feature distributions of real and fake images via Frechet distance between two Gaussians; it uses real data as reference and detects diversity issues, unlike IS which uses no real…
Mode collapse in GAN training
Generator produces few outputs ignoring data diversity, caused by chasing whatever fools the current discriminator; mitigate with minibatch discrimination, unrolled GANs, or Wasserstein loss.
Why ViTs need positional embeddings
Self-attention is permutation invariant so patch order is lost; positional embeddings restore spatial location. CNNs encode position implicitly via the fixed convolution grid.
Cross-attention for visual question answering
In cross-attention queries come from one modality and keys/values from the other, e.g. text queries attend over image features so the question selects relevant regions.
How Swin Transformer achieves linear attention
Swin computes attention within local non-overlapping windows of fixed size, making cost linear in patches, then shifts windows between layers so information crosses boundaries.
Inductive biases of ViT versus CNN
CNNs bake in locality and translation equivariance; a plain ViT has almost none beyond patch structure, so it must learn spatial relations from data, needing large datasets or strong pretraining.
Re-identification in multi-object tracking
Re-ID matches an object to its prior id using appearance embeddings, not just position; store track features and match re-entering detections by embedding similarity.
3D CNNs vs two-stream action recognition
3D CNNs learn spatiotemporal filters end to end but are heavy; two-stream splits RGB appearance and precomputed optical flow, strong but costly to compute flow.
Kalman filter for bounding-box tracking
State, transition, measurement models, and process plus measurement noise; predict then correct each frame. State holds box position and velocity; measurement is the detected box.
Brightness constancy and small-motion assumptions
Brightness constancy says a point's intensity is invariant under motion; small motion lets you linearize via Taylor expansion.
How to improve coarse segmentation boundaries?
Skip connections and higher-resolution features, boundary-aware losses, and point-based or CRF refinement.
How does Mask R-CNN do instance segmentation?
Faster R-CNN backbone plus RPN, then RoIAlign and a parallel mask head predicting per-class binary masks.
U-Net architecture and its skip connections
U-Net has a contracting encoder, an expanding decoder, and skip connections that concatenate matching-resolution encoder features into the decoder to recover spatial detail lost in downsampling.
Uses of the 1x1 convolution
A 1x1 conv is a per-pixel linear combination across channels; it reshapes channel depth cheaply and adds nonlinearity. Uses: dimensionality reduction in bottlenecks and channel mixing.
Receptive fields in convolutional networks
Receptive field is the input region affecting a neuron; it grows with depth, larger kernels, and stride. It matters for capturing context in detection and segmentation.
ResNet residual blocks and the degradation problem
A residual block learns F(x) and adds the identity input x, so layers fit a residual; this eases gradient flow and solves the degradation problem.
Regularization techniques for an overfitting CNN
Data augmentation expands the effective dataset, dropout prevents co-adaptation, weight decay penalizes large weights, plus early stopping and batchnorm.
The PnP problem in Structure from Motion
PnP recovers a camera's pose from known 3D points and their 2D projections; it registers new frames against the existing point cloud in SfM.
Incremental Structure from Motion pipeline
Detect and match features, estimate two-view geometry, triangulate, then incrementally add images with PnP and refine via bundle adjustment.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles