Deep Learning Practice · End Term · Sep 2025 · FN

**1. Identify the relevant counts:** Precision measures the fraction of detections that are correct. The detector produces 6 true positives and 4 false positives, so there are \(TP+FP=10\) predicted positives. **2. Divide correct detections by all detections:** \[\operatorname{Precision}=\frac{TP}{TP+FP}=\frac6{6+4}=0.6.\] Ground-truth objects that were missed affect recall rather than this denominator. **3. Interpret:** Sixty percent of the reported car detections are correct at the specified IoU threshold. **Answer: 0.6.**

**1. Identify the recall denominator:** Recall asks what fraction of actual objects were found. There are 10 ground-truth cars: 6 detected correctly and 4 missed. **2. Apply the formula:** \[\operatorname{Recall}=\frac{TP}{TP+FN}=\frac6{6+4}=0.6.\] **3. Interpret:** The detector finds 60% of actual cars. Recall happens to equal precision here because both the prediction count and ground-truth count are ten; the metrics need not usually be equal. **Answer: 0.6.**

**1. Reclassify the affected detection:** Increasing the IoU threshold makes one previous true positive fail the overlap requirement. Thus \(TP\) drops from six to five and \(FP\) rises from four to five. **2. Preserve the prediction count:** No detection is added or removed; only its evaluation label changes. Consequently \(TP+FP=5+5=10\). **3. Recompute precision:** \(5/10=0.50\), correct to two decimals. The associated ground truth also becomes missed, but false negatives do not enter the precision formula. **Answer: 0.50.**

**1. Count output positions:** A stride-one 1×1 convolution preserves the 28×28 spatial grid, giving \(28(28)=784\) output locations per filter. **2. Count work per output scalar:** One filter combines all 192 input channels, requiring 192 multiplications. With 64 filters, each spatial position needs \(192(64)=12{,}288\) multiplications. **3. Calculate and convert:** \[784(12{,}288)=9{,}633{,}792\text{ multiplications}=9.633792\text{ million}.\] This counts multiplications only, without adding additions or treating a multiply–accumulate as two operations. **Answer: 9.633792.** **Convention note:** The usual stride-one, spatially preserving 1×1 convolution is intended; no alternative stride is specified.

**1. Use rectified stereo geometry:** Disparity and depth satisfy \(d=fB/Z\), so \(Z=fB/d\). Focal length and disparity must use the same image-coordinate units. **2. Substitute:** \(f=100\) pixels, \(B=0.5\) metres and \(d=10\) pixels give \(Z=100(0.5)/10=5\). **3. Check units:** Pixels cancel between focal length and disparity, leaving metres. Greater disparity would correspond to a closer point for the same camera setup. **Answer: 5.**

**1. Count box quantities per cell:** In YOLO v1, each box contributes four coordinates and one confidence score. Two boxes therefore contribute \(2(5)=10\) channels. **2. Add cell-level class probabilities:** Each grid cell also predicts C=20 class probabilities. In v1 these class probabilities are shared across that cell’s box predictions, so they are added once, not once per box. **3. Form the tensor shape:** The output depth is \(5B+C=10+20=30\), and the complete grid output is \(7\times7\times30\). The question asks for depth, not all 1470 scalar entries. **Answer: 30.**

**1. Count one filter:** An 11×11 filter extending across three colour channels has \(11(11)(3)=363\) weights. **2. Count all filters:** With 96 independently learned filters, the layer has \(363(96)=34{,}848\) weights. **3. Apply the stated exclusion:** No biases are added. Each weight is reused at many spatial locations, so the number of output positions does not multiply the parameter count. **Answer: 34848.**

**1. Explain the training signal:** A stereo pair supplies two views of the same scene. Predicted disparity specifies where a pixel in one view should correspond in the other, permitting differentiable image reconstruction. **2. Explain the losses:** Photometric mismatch between the reconstructed view and observed target provides supervision without labelled depth maps. Left–right consistency adds a constraint that the two disparity predictions agree geometrically; smoothness can also be included. **3. Distinguish training from inference:** Stereo images provide the training constraint, while the trained network can predict disparity/depth information from one image at test time. **Answer: A, B, C — The model is trained using stereo image pairs but requires only a single image ; It uses the predicted disparity to warp the left image to match the right image ; The loss function is based on the photometric reconstruction error between .** A: **Correct:** Stereo pairs provide self-supervision in training, but the learned predictor takes one image at inference. B: **Correct:** Disparity supplies the correspondence used to warp one stereo image into the other. C: **Correct:** Photometric reconstruction is a central loss component, supplemented by consistency and regularity terms. D: **Incorrect:** The method avoids requiring ground-truth depth maps for training; LiDAR is not necessary for its stated supervision.

**1. Identify R-CNN’s repeated work:** Original R-CNN processes many cropped region proposals through the convolutional network separately. Overlapping proposals repeat expensive feature computation. **2. Explain feature sharing in Fast R-CNN:** The entire image is passed through the convolutional backbone once. Each proposed region then selects an area of that shared feature map. **3. Explain RoI pooling:** RoI pooling converts differently sized selected regions into fixed-size features for classification and box regression. Fast R-CNN still relies on proposals; eliminating external proposal generation is a different development. **Answer: A, B — It processes the whole image with a CNN once, rather than passing each ; It uses RoI Pooling to extract features for each region from the shared feature .** A: **Correct:** A single image-level backbone computation avoids repeatedly processing each region crop. B: **Correct:** RoI pooling obtains region-specific fixed-size features from the shared map. C: **Incorrect:** Fast R-CNN still consumes region proposals; it does not eliminate them. D: **Incorrect:** Fast R-CNN is proposal-based and is not the YOLO-style single-stage formulation.

**1. Explain overlap and matching:** \(\operatorname{IoU}=\text{intersection area}/\text{union area}\) compares a predicted box with a ground-truth box. A true positive also needs the appropriate class and an eligible one-to-one match. **2. Explain threshold failures:** A detection that fails the required IoU overlap is a false positive under the ordinary evaluation setting. A matching threshold is an evaluation rule, not a count of how many objects exist. **3. Explain AP and mAP:** Sorting predictions by confidence gives precision–recall operating points. Average precision summarizes the precision–recall curve according to the benchmark’s interpolation rule; mAP averages AP over classes, and some benchmarks also average over IoU thresholds. **Answer: A, B, C — IoU (Intersection over Union) measures the overlap between the predicted box ; A detection is considered a False Positive if the IoU is below a certain threshold ; mAP (Mean Average Precision) is calculated by averaging the Area Under the .** **Convention note:** Difficult, crowd or ignored annotations can receive special treatment in particular benchmarks. The question uses the ordinary matching rules and the usual precision–recall interpretation of AP. A: **Correct:** IoU quantifies the geometric overlap of the predicted and reference regions. B: **Correct:** Without an eligible ground-truth match meeting the threshold, a detection is ordinarily counted as a false positive. C: **Correct:** This is the intended AP-per-class then mean-across-classes description; precise interpolation depends on the benchmark. D: **Incorrect:** Finding all actual objects describes high recall. Many false alarms reduce precision rather than imply it.

**1. Separate distortion from perceptual realism:** Pixel-only MSE encourages an average of plausible high-resolution outputs, which can score well on PSNR yet appear overly smooth. **2. Explain the content term:** SRGAN compares feature representations from a pretrained VGG network, encouraging reconstructed content and perceptually meaningful structure. **3. Explain the adversarial term:** A discriminator encourages generated images to resemble natural high-resolution images. The perceptual/content and adversarial objectives work together; using VGG features does not turn the task into ImageNet classification. **Answer: A, B — Adversarial Loss: To encourage the generator to produce solutions on the ; Content (Perceptual) Loss: Based on feature maps from a pre-trained VGG .** A: **Correct:** The adversarial objective encourages realistic high-resolution image statistics. B: **Correct:** VGG-feature differences form the perceptual content loss in the standard SRGAN formulation. C: **Incorrect:** The word “only” excludes the adversarial and perceptual objectives central to the photo-realistic SRGAN approach. D: **Incorrect:** VGG is used to compute feature-based content loss, not to require a 1000-class image-label objective.

**1. Explain the global component:** A coarse network or encoder integrates information over a large receptive field. That helps infer scene layout and semantic depth relationships, though downsampling can blur boundaries. **2. Explain refinement:** A fine network or decoder recovers more spatially detailed predictions using global features and local information. Upsampling, rather than ordinary max pooling, increases resolution. **3. Explain skip connections:** Encoder features can be passed to later decoding stages to retain spatial detail lost during downsampling. These roles describe the architectures at a high level; their individual implementations differ. **Answer: A, B, D — The ”Coarse” network (or encoder) captures global scene structure but may ; The ”Fine” network (or decoder) refines local details.; Skip connections are often used to preserve spatial information lost during .** A: **Correct:** A coarse/global representation captures overall scene structure but can miss fine boundaries. B: **Correct:** Fine or decoder stages refine spatial detail in the prediction. C: **Incorrect:** Ordinary max pooling reduces or preserves resolution, rather than serving as the resolution-increasing operation. D: **Correct:** Skip connections help recover local information by making earlier spatial features available later.

**1. State YOLO’s formulation:** Original YOLO predicts box coordinates, confidence and class information in one image-level regression pipeline rather than first classifying many independent proposal crops. **2. Explain the grid:** The image is divided into an S×S grid, and each cell predicts its designated output quantities. The cell containing an object’s centre is responsible for that object in v1. **3. Explain context:** Processing the full image exposes the model to surrounding scene information. A Region Proposal Network belongs to Faster R-CNN, not this YOLO v1 pipeline. **Answer: A, B, D — It formulates object detection as a single regression problem, predicting ; It sees the entire image during training and testing, encoding contextual ; It splits the input image into an S × S grid..** **Convention note:** The question refers to the original grid-based YOLO description. Later YOLO versions alter prediction heads and box parameterizations. A: **Correct:** Joint regression of localization and class information is the original YOLO formulation. B: **Correct:** The full-image network has access to broader visual context during prediction. C: **Incorrect:** An RPN is a proposal mechanism associated with Faster R-CNN rather than YOLO v1. D: **Correct:** The S×S grid is a defining component of the original formulation.

**1. Identify degradation:** Simply making a plain network deeper can increase training error because optimization becomes more difficult. Residual learning addresses this optimization problem, which is distinct from merely reducing overfitting. **2. Explain a residual block:** Instead of learning only a direct mapping, a block computes \(y=F(x)+x\) when dimensions match. The identity route gives both activations and gradients an additional direct path. **3. Compare the intended architectures:** Standard ResNet designs can achieve strong or better accuracy with far fewer parameters than standard VGG designs, whose large fully connected layers are expensive. This is an empirical architectural comparison, not a theorem that every residual network must be smaller. **Answer: A, B, D — It solves the degradation problem where deeper networks have higher ; It uses ”skip connections” (or identity shortcuts) to allow gradients to flow ; It requires significantly less parameters than VGGNet to achieve similar .** A: **Correct:** Residual learning was introduced to address the increase in training error encountered when plain networks become deeper. B: **Correct:** Identity shortcuts provide direct paths that ease optimization and gradient propagation. C: **Incorrect:** The standard ResNet/VGG comparison does not require ResNet to have significantly more parameters. D: **Correct:** Canonical ResNets achieve competitive or better accuracy with substantially fewer parameters than canonical VGG networks.

**1. Explain local connectivity:** A convolutional filter reads a limited spatial neighbourhood instead of connecting each output unit to every image pixel. This uses the local structure of images. **2. Explain sharing:** The same filter is applied across positions. It reduces parameter count and lets the learned local feature be detected wherever it appears. **3. Clarify the translation claim:** Shared convolutions are primarily translation-equivariant: moving an input feature moves its feature-map response. Pooling or global aggregation can provide some position tolerance or invariance. This is the intended practical advantage behind the option’s wording. **Answer: A, B, C — Parameter Sharing: Weights are shared across the image, reducing the total ; Local Connectivity: Neurons are connected only to a local region of the input.; Translation Invariance: Features can be detected regardless of their position in .** **Convention note:** Select C in the question’s loose sense of position-independent feature detection. A convolution alone is translation-equivariant, not necessarily invariant; boundaries, strides and padding can further affect this property. A: **Correct:** Sharing one filter across positions uses fewer learned weights than separate position-specific dense connections. B: **Correct:** Local receptive fields match the spatial locality of image patterns. C: **Correct:** Under the intended position-tolerant feature-detection meaning, shared filters recognize a pattern at different positions; strict full-network translation invariance is not automatic. D: **Incorrect:** Connecting every input pixel to every hidden unit describes a dense layer rather than the local-connectivity advantage.

**1. Identify the random source:** Even at constant illumination, the number of photons arriving within an exposure fluctuates. The randomness arises from discrete photon detection events. **2. Match the noise model:** Photon shot noise is commonly modeled using a Poisson count, with variance equal to its mean in photon/electron-count units. It therefore depends on the signal level. **3. Separate electronics from arrivals:** Digitization and sensor readout introduce other noise sources, but the question specifically describes fluctuations in the photon count. **Answer: B — Photon shot noise.** A: **Incorrect:** Quantization noise arises from representing an analogue signal using discrete digital levels. B: **Correct:** Photon shot noise is the fluctuation in discrete photon arrivals or detected photoelectrons. C: **Incorrect:** Read noise comes from sensor readout electronics rather than the photon-arrival count. D: **Incorrect:** Salt-and-pepper noise describes isolated impulse corruptions, not the stated photon statistics.

**1. Identify the expensive dimension:** A K×K convolution from \(C_{in}\) channels to \(C_{out}\) channels needs \(K^2C_{in}C_{out}\) weights and corresponding channel-wise multiplication work at each position. **2. Insert a bottleneck:** A 1×1 convolution can map a large channel dimension to a smaller one before a 3×3 or 5×5 convolution. Reducing the latter layer’s input depth reduces its computation substantially. **3. Distinguish spatial and channel roles:** The bottleneck mixes channels without inherently increasing width or height. An activation after it can add nonlinearity, but depth reduction is the main motivation asked here. **Answer: A — To reduce the dimensionality (depth) of feature maps before expensive 3 × 3 .** A: **Correct:** Reducing channels before larger kernels lowers the expensive layers’ parameter and computation costs. B: **Incorrect:** A stride-one 1×1 convolution does not enlarge spatial dimensions. C: **Incorrect:** An added activation can introduce nonlinearity, but the bottleneck’s defining purpose here is reducing channel dimension. D: **Incorrect:** Global average pooling averages spatial values and is a different operation.

**1. Apply the assignment rule:** YOLO v1 assigns an object to the grid cell containing that object’s centre. This defines which cell predicts that object’s class and localization information. **2. Separate cell assignment from predictor selection:** Within the responsible cell, a particular box predictor is trained for the object according to its matching rule. That does not transfer responsibility to neighbouring cells. **3. Conclude:** The containing cell is responsible. Original YOLO v1 does not use the later anchor-box scheme described by the distractor. **Answer: C — That specific grid cell..** A: **Incorrect:** Responsibility is determined by the centre-containing cell, not neighbouring cells. B: **Incorrect:** Original YOLO v1 does not assign the object to an anchor box with the lowest IoU. C: **Correct:** The grid cell containing the object’s centre is responsible under the v1 rule. D: **Incorrect:** A network layer implements predictions, but the question asks about the grid assignment rule.

**1. Identify single-image cues:** Texture gradients, perspective and familiar object size can suggest relative or estimated depth within a single image. Their interpretation relies on assumptions or learned scene knowledge. **2. Identify the missing measurement:** Stereo parallax is the displacement of a scene point between two viewpoints. A single static image alone does not supply that two-view correspondence. **3. Connect to ambiguity:** Different 3D scenes can produce the same 2D projection, making metric depth underdetermined without additional assumptions or information. Missing stereo parallax is the relevant unavailable cue. **Answer: C — Stereo Parallax.** A: **Incorrect:** Texture changes across a surface can be observed in a single image and used as a depth cue. B: **Incorrect:** Familiar size and relative image size can supply monocular depth information, though not an unambiguous metric measurement alone. C: **Correct:** Stereo parallax requires observations from different viewpoints and is absent from a single static image. D: **Incorrect:** Converging lines and vanishing points are visible monocular perspective cues.

**1. Locate the RPN:** Faster R-CNN applies a proposal head to shared convolutional features. At feature-map locations it evaluates candidate anchor regions. **2. Identify its outputs:** For candidate regions, it predicts objectness scores and coordinate offsets that generate refined proposal boxes. These proposals are then passed to the detector’s downstream processing. **3. Separate proposals from final tasks:** The RPN identifies promising regions; the later detection head supplies the final category-specific predictions. It is not an evaluation-metric calculator. **Answer: B — To map sliding windows to object/not-object scores and bounding box .** A: **Incorrect:** The RPN predicts objectness rather than classifying proposals into all ImageNet categories. B: **Correct:** Object/not-object scores and box-coordinate predictions are the RPN’s central outputs. C: **Incorrect:** NMS can be used while filtering proposals, but it is not the RPN’s primary learned role or the complete final-output task. D: **Incorrect:** mAP is an evaluation metric, not the proposal head’s output.

**1. Start with one input position:** The receptive-field side starts at one. A stride-one 3×3 convolution expands it by \(3-1=2\) positions. **2. Expand layer by layer:** After the first layer the receptive field is 3×3; after the second it is 5×5; after the third it is 7×7. **3. Verify using the formula:** With stride one and no dilation, \(r=1+3(3-1)=7\). Intermediate nonlinearities affect the function learned, not this theoretical receptive-field size. **Answer: C — 7 x 7.** A: **Incorrect:** 3×3 is the receptive field after only one of the three layers. B: **Incorrect:** 5×5 is the receptive field after two layers. C: **Correct:** Three stride-one 3×3 layers produce a 7×7 theoretical receptive field. D: **Incorrect:** Adding the kernel sizes directly gives nine incorrectly; adjacent receptive fields overlap.

**1. Identify the multi-stage challenge:** Repeated restoration stages can lose information if each stage relies only on its immediate degraded/restored image representation. Useful features learned earlier should remain accessible. **2. Explain CSFF:** Cross-stage feature fusion transfers and combines feature representations from an earlier stage with features of the next stage. This strengthens information flow across successive restoration stages. **3. State the purpose:** The module reduces information loss and supports progressive refinement. It is a feature-communication component rather than a VGG loss computation or a random-noise generator. **Answer: B — To pass features from one stage to the next to minimize information loss..** A: **Incorrect:** Its central role is cross-stage information transfer, not reducing image dimensions for memory savings. B: **Correct:** Passing and fusing earlier-stage features helps preserve information for later refinement. C: **Incorrect:** Perceptual loss computation is not the CSFF module’s function. D: **Incorrect:** MPRNet’s restoration fusion module is not a GAN noise source.