Convolutional Networks

With a thousand background candidates per object, the summed training loss is dominated by easy background and the model drifts toward predicting nothing. Fixes constrain which candidates contribute: a fixed positive-to-negative sampling ratio, hard-negative mining, and an ignore band for ambiguous candidates.

mAP at IoU 0.5 forgives loose boxes, while averaging over the ten thresholds from 0.50 to 0.95 in 0.05 steps scores localisation quality directly. A model that classifies confidently but draws sloppy boxes tops the first and collapses on the second.

A head on the backbone's deepest feature map sees a stride near 32 pixels, so a 16-pixel sign occupies under one cell. Downsampling destroys its detail, and often no location is assigned to it. A feature pyramid fixes this.

Sharing asserts that the same local pattern means the same thing everywhere along an axis. It fails where position itself carries meaning: a spectrogram's frequency axis, absolute-coordinate targets, and tabular columns with no order.

Pixel accuracy is dominated by whichever classes own most of the pixels, while mIoU averages per-class intersection over union with every class weighted equally. Rare classes scoring near zero barely move accuracy but drag the mean down hard.

300 is not divisible by 32. The floor in the output-shape formula drops a row at every odd-sized stage, so the encoder produces 37 where the decoder's doubling produces 36, and concatenation needs identical height and width.

Budget the worst case, not the average. A one-stage head costs the same on every frame; a two-stage head's second pass scales with proposal count, so busy scenes are its slowest. Decide on tail latency first, then on the accuracy you need.

Fine spatial detail. A stride-4 first layer samples the image on a coarse grid, aliasing away structure finer than four pixels, and one wide linear filter plus one nonlinearity is a weaker map than a stride-1 3x3 stack.

A depthwise stage removes arithmetic but not data movement: it streams the full activation tensor while doing only k-squared multiply-accumulates per element. Runtime becomes memory-bound, and one layer becoming two adds further passes.

A panoptic output allows one label per pixel, so conflicts follow a fixed precedence: instance masks are painted in confidence order, each keeping only unclaimed pixels, and stuff fills the remainder. Fragments left too small are dropped.

Dilation spreads a kernel's taps apart, covering a wider region with no extra weights and no downsampling. The costs are sparse sampling between the taps, full-resolution activation memory, and gridding when layers repeat one rate.

Replace it when the downsample itself should be learned: a stride-2 convolution trains a weighted summary per window instead of applying a fixed maximum. It costs parameters, multiply-adds and a useful prior, and pays off most when data is plentiful.

Concatenation keeps every earlier feature map intact so later layers can select among them; addition merges them irreversibly but holds channel count fixed. Choose concatenation for feature reuse at moderate depth, addition for very deep throughput-sensitive backbones.

A transposed convolution writes a scaled copy of its kernel into the output at stride spacing. When the kernel size is not divisible by the stride, some output positions receive more overlapping writes than others, and that uneven overlap prints a periodic checkerboard.

A 3D kernel spans time, height and width at once, so weights and compute scale with clip length too. The (2+1)D form splits it into a spatial convolution then a temporal one: fewer weights, an extra nonlinearity, easier optimization.

With 12 instances, recall moves in steps of about 8 points, so that class's AP is a coarse, high-variance statistic. Publish per-class AP beside its instance count with an uncertainty estimate, and never let the headline mean carry a ship decision it cannot support.