Homework 4

starter code dense data colab notebook

In this homework, we will implement an object detector from scratch.

This assignment should be solved individually. No collaboration, sharing of solutions, or exchange of models is allowed. Please, do not directly copy existing code from anywhere other than your previous solutions, or the previous master solution. We will check assignments for duplicates. See below for more details.

The goal of our object detector is to find karts, bombs/projectiles, and pickup items. We’ll use the same dataset and a similar starter code to homework 4. You’ll also need to employ similar data augmentation to make detection work.

box

We will implement a point-based object detector. Your detector will represent each object by a single point at its center. This point is easy to detect by a dense prediction network, similar to HW3.

Point-based object detection

You’ll take your segmentation network and repurpose it for object detection. Specifically, you’ll predict a dense heatmap of object centers, as shown below:

heatmap

Each “peak” (local maxima) in this heatmap corresponds to a detected object. We start by implementing a function that extracts all local maxima (peaks) from a given heatmap.

Peak extraction

A peak in a heatmap is any point that is

Implement the extract_peak function in models.py to extract these local maxima. The input to extract_peak is a 2d heatmap, the output is a list of peaks with score, x and y location of the center of the peak. The score is the value of the heatmap at the peak, and is used to rank detections later. A peak with a high score is a more important detection.

Test your peak extractor using

python -m grader homework

Hint: There may be many peaks close to an object center. Only return local maxima (points that have a higher response that any point in their surrounding).

Hint: Implement local maxima detection using max pooling.

Hint: Max pooling expects an order-4 tensor as an input, use heatmap[None,None] to convert the input heatmap to an order-4 tensor.

Relevant Operations

Next, we’ll implement the detector. This involves several steps before you’ll be able to grade any part of it: First, you need to design your detection model. Second, you’ll need to train your model to predict the above heatmap. Third, you’ll need to use extract_peak to extract objects. We only grade the final detections and no intermediate output.

CNN model

You model should take an image as an input and predict a heatmap for each object. Each heatmap is an independent 0-1 output map. The heatmap is zero almost everywhere, except at object centers. Use a model similar to your solution to HW3 or the master solution. Implement your model in Detector in models.py.

Hint: Use the sigmoid and BCEWithLogitsLoss

Hint: Implement __init__ and forward first, worry about detect and detect_with_size later.

Training your CNN

The training code for your detector in train.py will look quite similar to HW3. DetectionSuperTuxDataset loads the dataset. Unlike HW3, the annotations for this homework are not yet in a format that can be batched by pytorch. You can convert the detection annotations to a heatmap by calling the dense_transforms.ToHeatmap transform. ToHeatmap changes just the label and leaves the image unchanged. Call ToTensor to change the image. dense_transforms.py also contains several other annotations that you might find useful. utils.py shows how to use some of those transformations.

python -m homework.utils

It is highly recommended that you log your training using tensorboard.

Hint: Similar to HW3 the positive and negative samples are highly imbalanced. Try using the focal loss.

Relevant Operations

Object detection

The final step of your detector extracts local maxima from each predicted heatmap. Each local maxima corresponds to a positive detection. The function detect returns those detections as a tuple of four numbers: The object class (0..2), the confidence of the detection (float, higher means more confident, see evaluation), the x and y location of the object center (center of the predicted bounding box). Your detection function may return up to 100 detections per image, each detection comes with a confidence. You’ll pay a higher price for getting a high confidence detection wrong. The value of the heatmap at the local maxima (peak score) is a good confidence measure. Use the extract_peak function to find detected objects.

Grading

We will evaluate your detector using average precision (AP). See here or here for a good explanation for average precision.

Average precision sorts all your detections (over the entire dataset) by their confidence. It then measures the precision (true positive / number of predictions) and recall (true positive / number of objects) for each class. Average precision is the mean precision at a recall of 0, 0.1, 0.2, …, 1. We use two different criteria to determine if a detection is positive or negative:

The dataset contains quite a few small objects (smaller than 20px area). We ignore all small objects and do not penalize any predictions close to small objects.

Call the grader to compute your average precision

python -m grader homework

We will grade your assignment using the following rubric:

Warning: The validation set is significantly easier than the test set. Our master solution reaches a 0.820 / 0.609 / 0.894 validation AP on the point inside box metric, but only 0.741 / 0.531 / 0.887 test AP. To be on the save side make sure your solution surpasses the required AP by 0.1 .

Submission

Once you finished the assignment, create a submission bundle using

python bundle.py homework [YOUR UT ID]

and submit the zip file online. If you want to double-check that your zip file was properly created, you can grade it again

python -m grader [YOUR UT ID].zip

Extra credit (9pts): Full object detection

You can earn some extra credit by implementing a full-fledged object detector. To do so you’ll need to predict the size of the object in question. The easiest way to do this is to predict two more output channels, in addition to the 3-channel heatmap. We will evaluate AP at an intersection over union (overlap) of 0.5.

Hint: Make sure, to only apply the size loss at object centers.

Hint: Carefully balance the size and peak losses.

Running your assignment on google colab

You might need a GPU to train your models. You can get a free one on google colab. We provide you with a ipython notebook that can get you started on colab for each homework. Follow the instructions below to use it.

Honor code

This assignment should be solved individually.

What interaction with classmates is allowed?

What interaction is not allowed?

Ways students failed in past years (do not do this):

Installation and setup

Installing python 3

Go to https://www.python.org/downloads/ to download python 3. Alternatively, you can install a python distribution such as Anaconda. Please select python 3 (not python 2).

Installing the dependencies

Install all dependencies using

pip install -r requirements.txt

Note: On some systems, you might be required to use pip3 instead of pip for python 3.

If you’re using conda use

conda env create environment.yml

Manual installation of pytorch

Go to https://pytorch.org/get-started/locally/ then select the stable Pytorch build, your OS, package (pip if you installed python 3 directly, conda if you installed Anaconda), python version, cuda version. Run the provided command. Note that cuda is not required, you can select cuda = None if you don’t have a GPU or don’t want to do GPU training locally. We will provide instruction for doing remote GPU training on Google Colab for free.

Manual installation of the Python Imaging Library (PIL)

The easiest way to install the PIL is through pip/pip3 or conda.

pip install -U Pillow

There are a few important considerations when using PIL. First, make sure that your OS uses libjpeg-turbo and not the slower libjpeg (all modern Ubuntu versions do by default). Second, if you’re frustrated with slow image transformations in PIL use Pillow-SIMD instead:

CC="cc -mavx2" pip install -U --force-reinstall Pillow-SIMD

The CC="cc -mavx2" is only needed if your CPU supports AVX2 instructions. pip will most likely complain a bit about missing dependencies. Install them, either through conda, or your favorite package manager (apt, brew, …).